Commit graph

3,718 commits

Author SHA1 Message Date
Julian Knutsen
f3c47390e7
fix(storage): refuse graph_mode workspaces on library opens (#7721)
#7671 fences graph workspaces in cmd/bd: command admission, the store
factories and the other CLI paths refuse one, and #7722 extends that
fence to a .beads holding the graph-preview-format marker whatever its
metadata says. But the public library opens (beads.OpenBestAvailable,
OpenBestAvailableWith, OpenFromConfig, OpenGated) and the Dolt store's
metadata-driven constructors (NewFromConfig, NewFromConfigWithOptions,
NewFromConfigWithCLIOptions) still opened either kind of graph workspace
as an Issue store and ran Issue schema migrations into the graph
database (or, with no metadata beside the marker, created an Issue
database inside the graph workspace).

- Move the typed refusal into internal/configfile
  (UnsupportedWorkspaceFormatError with BeadsDir, GraphMode and
  FormatMarker; CheckWorkspaceFormat; GraphFormatMarkerFile;
  HasGraphFormatMarker) with byte-identical text and hints for both the
  graph_mode and the marker refusal, and move cmd/bd's discovery
  pre-checks there as LoadAndCheckWorkspaceFormat and
  CheckDiscoveredWorkspaceFormat (LoadForDiscovery, which never migrates
  a legacy config.json; unreadable metadata beside the marker is refused
  as the marker's format). cmd/bd's checkWorkspaceFormat,
  loadAndCheckWorkspaceFormat, hasGraphFormatMarker and
  checkDiscoveredWorkspaceFormat become thin delegates, so the CLI keeps
  its behaviour (exit 16 with the same hints) and every caller of them is
  unchanged; bootstrap's ancestor walk builds the configfile type.
- Every public metadata-driven open runs CheckDiscoveredWorkspaceFormat
  as its first statement: before OpenGated takes its gates and before
  OpenBestAvailableWith's Load can migrate config.json. A refused open
  leaves the workspace byte-identical. The type is exported as
  beads.UnsupportedWorkspaceFormatError.
- The Dolt constructors run CheckDiscoveredWorkspaceFormat before
  configfile.Load, so a graph workspace marked only by a legacy
  config.json, or by the marker beside stripped, missing or unparseable
  metadata, is refused before Load can migrate it into metadata.json
  (and drop graph_* keys this build does not know). requireDoltBackend
  re-checks the loaded config as a backstop, so they also refuse before
  resolving a server or writing a port file or Dolt root.

Tests: configfile unit tests for both refusals (exact marker text and
hint, graph_mode winning over the marker, unparseable metadata with and
without the marker); the public opens and the three Dolt constructors
against graph_mode workspaces and against marker-only, stripped
(metadata.json and legacy config.json), server, explicit dependency and
unparseable-metadata marker workspaces, each refused with the typed
error and the workspace left unchanged.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-20 09:29:54 +00:00
Julian Knutsen
1fcd715e75
fix(cli): refuse graph_mode workspaces and keep their metadata marker on save (#7671)
* fix(configfile): carry graph_mode and graph_* binding through Save

Graph-preview builds of bd persist a per-workspace format in
metadata.json: "graph_mode": "link" plus graph_scope_url,
graph_authority_id, graph_workspace, graph_schema_version and
graph_ready, beside an ordinary Dolt backend. configfile.Config had no
such fields, so every Save, and Load's config.json -> metadata.json
migration, rewrote the file without them, turning a graph workspace's
metadata into an ordinary Issue workspace's.

Config now carries the marker and the binding as omitempty fields, so
a Save round-trips them unchanged and metadata.json written for an
ordinary workspace is byte-for-byte what it was. GetGraphMode resolves
the marker: absent or "dependency" is the historical default, "link"
is recognized, and any other value is an error rather than a fallback
to the Issue store. Nothing in this build opens a non-default mode;
the CLI refusal follows in the next commit.

Ported from the donnabox fork commit deea8d57df (graph_mode.go,
graph_mode_test.go and the configfile.go fields), without any graph
code. The binding-survives-Save and omitted-from-ordinary-metadata
tests are new.

Co-authored-by: Donna Box <dbox@boxenllc.com>
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(cli): refuse graph_mode workspaces before opening them

A workspace whose metadata.json declares a graph_mode other than
"dependency" holds a graph database behind an ordinary Dolt backend,
and nothing else in its metadata says so. bd admitted it like any
current workspace: `bd list` ran the Issue schema initialization
against the graph database (beads_graph_<id>), `bd init --force` wrote
a .gitignore and a .beads/dolt root into it, `bd dolt start` started a
server for it, and `bd doctor` reported on it as an Issue store.

guardLegacyUpgradeWorkspace, the admission check every store-opening
command, store-less command with workspace effects, init, doctor and
bootstrap already run before touching a workspace, now refuses such a
workspace first, before the legacy classifications that would admit
it or report it as a legacy store to migrate. An unknown graph_mode
value is refused the same way. The refusal is a typed error naming the
workspace and the format, rendered with a hint (as JSON under --json)
and exit code 16, ExitUnsupportedWorkspaceFormat, so a wrapper can
tell "needs a different bd build" from a generic failure.
`bd version`, `bd --version`, `bd help` and completion stay exempt, as
they are for legacy workspaces. validateConfiguredBackend refuses the
same metadata, which covers the store-open paths that skip command
admission (routed writes, cross-repository reads, create --repo).

Tests: guard and store-factory unit tests; a subprocess e2e test that
show/list/create/context/where/dolt start/doctor/bootstrap/init are
refused with exit 16 and leave .beads byte-identical (metadata.json and
legacy config.json, link and unknown markers, embedded and server
layouts), and that store-less commands still run; an embedded test
that an ordinary workspace, with the marker absent or "dependency",
still creates and lists issues and keeps the marker through a
metadata.json rewrite.

The guard hunk and graph_mode_admission_e2e_test.go are ported from
the donnabox fork commit deea8d57df; the exit code, rendering, store
factory backstop and remaining tests are new.

Co-authored-by: Donna Box <dbox@boxenllc.com>
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(cli): refuse graph workspaces reached outside command admission

Command admission refuses a graph_mode workspace a command runs in, but
bd also reaches workspaces before or without admission: a create --repo
target, the path --db or BEADS_DB names, and the stores the factories
open for routed and cross-repository work. On those paths a workspace
marked only by a legacy config.json was changed before anything refused
it: configfile.Load migrates config.json into metadata.json before the
format check sees the loaded metadata, and create --repo initialized
such a target as a new workspace. Where the refusal did come, create
reported it with exit 1 rather than 16.

- checkDiscoveredWorkspaceFormat checks a workspace's metadata as
  configfile.LoadForDiscovery reads it, which never rewrites it.
  ensureBeadsDirForPath, the dry-run target open and the four store
  factories (cgo and nocgo builds) run it before their own load.
- resolveBeadsDirForDBPath and findDatabaseInBeadsDir locate a
  workspace with LoadForDiscovery, so resolving --db, BEADS_DB or a
  create target no longer migrates it.
- bd create renders a refused target as the workspace refusal, exit 16
  with its hint, on the initializing and both dry-run paths. A dry run
  without --parent opens no target, so it now checks a local one before
  rendering the preview (checkRepoTargetFormat).
- bd bootstrap, when discovery finds no workspace, reads each .beads
  from the working directory up to the repository root without
  migrating it and refuses one in a format this build cannot open, the
  way commands without a workspace already do. A .beads marked only by
  config.json is passed over by discovery, so bootstrap reported "no
  workspace" there instead of exit 16. The walk now returns its
  candidates (undiscoveredWorkspaceCandidates) to both checks: handing
  it the check as a func(string) error made gosec's G702 taint analysis
  connect BEADS_DIR to every func(string) error that runs git.
- backendNameForErrorFraming reads with LoadForDiscovery, and
  loadDoltBackendConfig, bd dolt status and bd dolt show refuse a graph
  workspace before their migrating configfile.Load. Admission refuses
  first on every path that reaches them today; this keeps a new caller
  from migrating a legacy config.json before the refusal.
- The CHANGELOG entry and the configfile comment now claim what holds:
  a routed read fails without touching a graph workspace but not with
  code 16, and a Save keeps only the known graph_* fields, with the
  tags graph-preview builds declare. workspace_gate.go no longer lists
  the pre-gate discovery reads as migrating.
- New e2e rows run create --repo from a workspace and from outside one,
  create --dry-run with and without --parent, --db, BEADS_DB, init
  --proxied-server, bootstrap in a workspace nested in a graph one, and
  the hook shim's bd hooks run pre-commit against every fixture; each
  refusal must exit 16 and name the format and the hint. bootstrap in a
  workspace marked only by config.json now refuses like the other
  commands. The store factory test also runs on a config.json graph
  workspace, TestWorkspaceMetadataReadersLeaveGraphWorkspaceUnchanged
  covers the framing, bd dolt loader and diagnostic reads, and
  TestNocgoStoreFactoriesRefuseGraphWorkspace covers the CGO_ENABLED=0
  factories; bazel.yml's pure lane runs it by name (PURE_CMD_BD_TESTS,
  pinned in scripts/ci_workflow_test.go).

Measured with gocyclo/gocognit: every function this changes stays below
the PR's worst changed function (guardLegacyUpgradeWorkspaceForInit
29/28, resolveBeadsDirForDBPath 26/36); showDoltConfig is 19/31 (was
18/30) and guardUndiscoveredLegacyWorkspace 3/3 (was 9/12).

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Donna Box <dbox@boxenllc.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-20 00:05:39 +00:00
Julian Knutsen
2563f021ed
perf(httpclient): remember the server's external-dependency policy across processes (#7700)
* perf(httpclient): remember the server's external-dependency policy across processes

Every bd ready over an HTTP workspace fetched GET /v0/beads/context before
its listing. listReadyWork itself needs no handshake; the fetch came from the
external-dependency decorator asking ServerEnforcesExternalDependencyPolicy
whether the server already applies the policy. In a one-shot CLI that is one
extra round trip per process (about 54 ms at a 52 ms RTT) to relearn a bit
that only changes when the server is restarted with a different composition.

The store opened from a workspace now keeps that answer in the user cache
directory (beads/http-policy/), keyed by the server URL and the project the
workspace pinned at bd connect, for BEADS_HTTP_POLICY_CACHE_TTL (5 minutes by
default; off disables it). A fresh entry answers the probe without dialing; a
handshake this process already holds wins and refreshes the entry. A
workspace with no project pin, and a store an embedder opens through
OpenWith, are not cached. Writing the entry is best effort.

The staleness this allows is bounded by the TTL: a server restarted without
the policy within the window is still trusted to enforce it until the entry
ages out. Carrying the policy on every response would close that window but
needs a server change and does nothing for servers already deployed.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of gascity-agent
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(httpclient): claims and closes ask the server for the policy, never the cache

Review follow-up to the policy cache. The claim and close guards (direct
claim, claim-next, batch close, the lifecycle close and claim, and the legacy
ClaimReadyIssue and CloseIssueChecked) handshake anyway, and a stale cached
"the server enforces it" would stand down their client-side refusal of
externally blocked work. Those entry points now mark their context with
storage.WithLivePolicyProbe, and httpclient.Store answers such a probe from
the server, refreshing the cache entry. Reads keep the cached answer.

Tests: externaldeps TestClaimsAndClosesAskForALivePolicyProbe and
TestRemoteClaimsAndClosesAskForALivePolicyProbe pin which entry points probe
live (every claim and close) and which do not (ready listings), over a local
and a remote store; httpclient TestALiveProbeSkipsThePolicyCache pins the
store side; TestPolicyCacheRefusesEntriesItCannotTrust now also covers an
entry written for another server URL.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of gascity-agent
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-19 19:05:55 +00:00
Julian Knutsen
d28f56cd46
perf(startup): initialize packages on four Ps, not every CPU (#7701)
* perf(startup): initialize packages on four Ps, not every CPU

bd links several thousand packages whose init functions allocate about
20 MB before main runs, which triggers four or five garbage collections.
Each one starts mark workers and stops the world across every P, and on a
large host that costs more than the collection itself: on a 192-CPU machine
`bd version` spent about 20 of its 110 ms there, paid again by every bd
process a script or an orchestrator starts.

internal/initprocs lowers GOMAXPROCS to 4 for the duration of package
initialization. It imports only runtime and syscall, so it initializes
ahead of the heavy packages, and main hands GOMAXPROCS back first thing:
to the GOMAXPROCS environment value when one is set, else to the runtime
default, which keeps tracking the CPU limit.

cmd/bd's TestMain does the same hand-back, since tests never run main, and
resets the -test.parallel default testing derived from the capped count.

Measured on that host (40 runs each, `bd version`): p50 103 -> 86 ms.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of gascity-agent
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(initprocs): tell an explicit GOMAXPROCS hand-back from the runtime default

Review follow-up. TestDoneHonorsAnExplicitGOMAXPROCS swaps the two runtime
calls for recorders and pins that Done restores the exact value when
GOMAXPROCS is set and calls SetDefaultGOMAXPROCS when it is not; the earlier
test only saw the resulting count, which is the same either way on a host
whose default equals the saved value.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of gascity-agent
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-19 17:15:54 +00:00
Julian Knutsen
18b5c3cc39
fix(httpclient): name the connect failure when bd serve cannot be reached (#7702)
* fix(httpclient): name the connect failure when bd serve cannot be reached

A workspace whose server does not answer reported the transport's text:
`cannot reach bd serve at https://host:30699: Get "https://host:30699/v0/beads/context":
dial tcp 100.85.56.42:30699: i/o timeout`. It names the request rather than
the address that failed, and says nothing about what a dial timeout usually
means. An operator chasing an unreachable store had to know that "i/o
timeout" on a dial means the SYN went unanswered.

A request whose connection could not be opened now fails with a
*wire.DialError inside the *ConnectError: a refused connect names the
address and says nothing is listening there; a connect that times out names
the address that never answered, how long the attempt waited, and that a
firewall dropping the connection or a down host is the usual cause; a host
name that does not resolve is named as such. The transport's error stays
reachable through Unwrap, so errors.Is checks (ECONNREFUSED, timeouts) and
*ConnectError classification are unchanged, and a dial the caller canceled
keeps its own error. The events:watch connect gets the same classification.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of gascity-agent
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(httpclient): keep the transport's dial text in connect errors

Review follow-up. The refused and DNS messages dropped "connection refused"
and "dial tcp", the vocabulary gc's text classifier matches to tell an
unavailable store from other failures; without it the dispatcher's serve
loop exits and control beads are quarantined. Every DialError message now
ends with the transport's own *net.OpError text in brackets, after the
explanation.

TestACanceledDialIsTheCallersOwn now actually reaches the ctx.Err() guard:
its round tripper cancels and hands back a genuine refused dial, which the
guard alone keeps from becoming a DialError (removing the guard fails it).

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of gascity-agent
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-19 16:54:25 +00:00
Julian Knutsen
1de9b59324
fix(setup): bd setup claude --global manages the section in ~/.claude/CLAUDE.md, not the current directory (#7696)
* fix(setup): keep bd setup claude --global's section in ~/.claude/CLAUDE.md

bd setup claude --global put its hook in ~/.claude/settings.json but wrote
the managed beads section into the current directory's CLAUDE.md, and
--remove --global stripped it from there. Run from $HOME it created
~/CLAUDE.md with project scaffolding, which stops Claude Code reading
AGENTS.md in projects below it; run inside a project it removed the
section that project's own install had written, leaving its hooks.

A global install and removal now manage the section in Claude Code's
user-level ~/.claude/CLAUDE.md (created holding only the section, no
project scaffolding), with hints naming bd setup claude --global. With
global or plugin hooks, --check reports that file when the project's
instructions have no beads section; a project install is still judged
on the project's file alone. Project installs are unchanged.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(setup): render the user-level CLAUDE.md section independently of the project

~/.claude/CLAUDE.md is shared by every project, but bd setup claude --global
rendered and --check judged its section with the current directory's beads
config (sync.remote, sync.git-remote, no-push). From a project with a remote
and one without, --check called the other's install stale and its hint
flipped the file back. The user-level file now always renders with the
default options, so install and --check agree across projects.

Also pin --check's fallback to the project's file and project hint when
neither file has a section, tell existing --global users to re-run it, and
name ~/.claude/CLAUDE.md in the ide-setup recipe table.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(setup): judge the user-level Claude files from $HOME and ~/.claude

Run from the home directory, the project's .claude/settings.json is the
global one, so bd setup claude --check reported a global install as project
hooks and judged ~/CLAUDE.md; its "Run: bd setup claude" hint then created
the ~/CLAUDE.md a global install exists to avoid. Run from ~/.claude, the
project's CLAUDE.md is the user-level file, and --check judged it with that
directory's render options and the project hint.

--check now reports hooks as global from $HOME and judges the user-level
file there and from ~/.claude with the fixed options and the --global hint.
With user-level hooks and no section in either file it names --global
unless the project has its own CLAUDE.md. A --remove run from $HOME, which
removes the global hooks, also removes the user-level section.

The init agent-artifact test ran bd setup claude with HOME at the project,
so its checks would now see a home directory; it gets a separate HOME and
its goldens are unchanged.

Rebase note (maintainer): rebased onto main after #7678 merged; main's
agent-artifact harness already runs bd with a separate HOME, so this
commit's harness edits to cmd/bd/agent_artifacts_embedded_test.go were
dropped.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(setup): pin --check's user-level choice beside project instructions

With global hooks, --check judges ~/.claude/CLAUDE.md when the project's
own CLAUDE.md has no beads section but the user-level file has one, and
when run from the home directory whose ~/CLAUDE.md has no section. Nothing
exercised either half of that condition: replacing it with false left the
suite and goldens green. Add a subtest for each; each now fails when its
half of the condition is dropped.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(setup): name --global when --check finds no hooks from $HOME

From the home directory --check with no hooks said 'Run: bd setup claude'.
Following it runs a project install there, which creates ~/CLAUDE.md with
project scaffolding, the file that stops Claude Code reading AGENTS.md in
the projects below it. The missing-section hint at home already names
--global; make the missing-hooks hint match, with a subtest.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(setup): qualify the --global CHANGELOG entry; watch ~/.claude/CLAUDE.md

The CHANGELOG said --check from the home directory judges the user-level
files, but a beads section still in ~/CLAUDE.md is judged instead, and
the --global hint for a missing section applies only when the project has
no CLAUDE.md of its own. Say both, mention the new missing-hooks hint at
home, and re-wrap the paragraph.

The embedded agent-artifact harness comment said a global install writes
$HOME/CLAUDE.md; it now writes $HOME/.claude/CLAUDE.md. Track that path
too and reword the comment. The bd init goldens do not change.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(setup): word the user-level CLAUDE.md section for repositories that use beads

bd setup claude --global writes its section to ~/.claude/CLAUDE.md, which
Claude Code loads in every project, but rendered it with the project
wording: "This project uses bd (beads)" and "Use bd for ALL task tracking —
do NOT use TodoWrite". In a repository without beads that told the agent to
give up its own task tools and pointed it at bd, whose answer there is
"run 'bd init'", while the global hook, bd prime, stays silent outside a
workspace.

RenderOpts.UserLevel rewords the opening line and the task-tracking rules of
both profiles so they apply only in beads repositories. The global env fixes
it through agents.UserLevelRenderOpts, so install, in-place update and
--check agree on the hash from every directory, and a project-worded section
left in ~/.claude/CLAUDE.md reads as stale and is reworded by --global. The
new minimal-user and full-user variants are pinned by golden and hash; the
project rendering is unchanged.

Also pin sameDir's symlink following: --check from a symlink to $HOME or to
~/.claude still judges the user-level files.

Refreshed onto main past #7692, whose --remove takes the blank lines
around the section with it and whose setup appends a section after one
blank line: claude-setup-global-beside-project-install.golden no longer
records the old hole (an extra blank line before the appended section,
two left behind by --remove --global).

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-19 12:22:04 +00:00
Julian Knutsen
21dac62a4c
fix(init): drop bd dolt push from AGENTS.md Quick Reference without a remote (#7694)
A new AGENTS.md written by bd init rendered its managed section with the
HasRemote/NoPush rule, but wrote the template text around it verbatim, so
the Quick Reference still listed `bd dolt push` in workspaces with no Dolt
remote or declared no-push. Add EmbeddedDefaultWithOpts, sharing one
omitsDoltPush predicate with the managed-section renderer, and use it for
new AGENTS.md files. Managed-section bytes and hashes are unchanged; an
existing AGENTS.md and a custom --agents-template are written as before.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-19 10:36:59 +00:00
gascity-eddie[bot]
4c473b4580
fix(query): report an uppercase duration unit as an unparseable time, not a negated internal error (#7683)
* refactor(timeparsing): export IsCompactDuration

The query lexer keeps its own copy of the compact-duration unit set, and
that copy has drifted from the grammar: it accepts the uppercase units
H D W M Y, which ParseCompactDuration rejects. Export the classifier so
the lexer can ask timeparsing instead of keeping a copy, and pin that it
agrees with ParseCompactDuration on every row, uppercase spellings
included.

AC: AC-1
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(query): classify durations with timeparsing.IsCompactDuration

The query lexer kept its own copy of the compact-duration unit set:
isDurationSuffix for single runes and durationWords for "min". The copy
had drifted. It accepted H D W M Y, which the case-sensitive grammar
rejects, so "created>30M" reached the evaluator as a duration and failed
with the internal error `not a compact duration: "-30M"`, quoting a value
the user never typed.

Classify the whole digit-led lexeme with timeparsing.IsCompactDuration
and delete the copy (isDurationSuffix, durationWords, acceptDurationWord).
Uppercase units now lex as identifiers and fail as an unparseable time;
every lowercase input lexes as before. A unit added to compactDurationRe
now reaches the lexer with no edit here: with "sec" added to the regex
alone, TestCompactDurationGrammarReachesQueryPlane fails on the old lexer
and passes on this one. Retire that test's text naming the deleted copy.

AC: AC-2
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(query): pin that uppercase duration units lex as identifiers

"updated>30M", "updated>30H" and "label=30D" lex the value as one
TokenIdent with the text as typed; "updated>30m" and "updated>+30d"
still lex as TokenDuration. On the base lexer the three uppercase rows
fail (typed DURATION).

AC: AC-3
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(query): pin the error for an uppercase duration unit

created>, updated<, closed>= and started<= compared to 30M, 30D, 30W,
30Y, 30H or 1H, alone or in "... OR status=open", fail with
`cannot parse time expression: "<value as typed>"`, never with the
internal `not a compact duration` error and never quoting the negated
"-<value>". 30MIN and 30Min, which always took this path, are the
controls. On the base lexer 30 of the 40 subtests fail; the 10 controls
pass.

AC: AC-4
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(query): bind the lexer's duration grammar to timeparsing both ways

TestCompactDurationGrammarReachesQueryPlane asserted only that every
suffix ParseCompactDuration accepts lexes as a duration. Assert the
reverse too: a suffix it rejects must not lex as TokenDuration, or the
evaluator fails on it with ParseCompactDuration's internal error about a
negated value. Widen the candidates to every one- and two-letter ASCII
suffix in either case, every three-letter lowercase one, and every case
spelling of an accepted three-letter unit.

On the base lexer the reverse sweep fails on exactly [D H M W Y]. A
mutant that case-folds before classifying fails it on those plus the
seven other spellings of "min".

AC: AC-5
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(query): pin that a signed uppercase unit stops at the parser

"created>+30M" and "created>-30M" fail Parse and EvaluateAt with
`unexpected token "M" at position 11 (expected end of query)`, as
"created>+30x" already does: a signed digit run has no identifier
fallback, so "+30" is a number and "M" a stray token. On the base lexer
both parsed as a duration and failed in the evaluator with
`not a compact duration: "-30M"` and `"--30M"`; the +30x control passes
on both.

AC: AC-6
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(query): pin the values rejecting uppercase units leaves alone

At a fixed now, created>+30d is still 30 days ago, updated>5pm still
17:00 that day, and label=30M / title=30M still filter the literal
"30M". TestQueryMinutesUnitResolvesAgo already pins created>30m (30
months ago) and created>30min (30 minutes ago). These rows pass on the
base lexer as well, which is the point; a lexer mutant that stops typing
signed lexemes as durations fails the +30d row.

AC: AC-7
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 09:18:13 +00:00
gascity-eddie[bot]
994f5b82dd
fix(config): stop bd config set truncating config.yaml at a 64 KiB line (#7670)
* fix(config): stop bd config set truncating config.yaml at a 64 KiB line

updateYamlKey's flat-key path reads config.yaml through a default
bufio.Scanner and never checked scanner.Err(). A line of
bufio.MaxScanTokenSize (65,536) bytes or more makes Scan return false
partway through the document, so the function returned the lines read
so far with a nil error and setYamlConfigAtPath wrote them over the
file, dropping the long line and everything after it. Measured at
5c9c02ab77: a 65,550-byte document came back as the 8-byte
`k1: "zz"` when the key was present, and as 16 bytes when it was absent
(the new key appended onto the truncated prefix).

Check scanner.Err() after the loop and refuse with an error that wraps
bufio.ErrTooLong. It is returned before setYamlConfigAtPath's
os.WriteFile, so `bd config set` and `set-many` now fail and leave the
file as it was. The guard returns only on a scan error, so a document
whose lines all fit gets the same rewrite as before. Unset
(commentOutYamlKey) already splits instead of scanning, and dotted keys
rewrite through yaml.Node, so neither needs it.

TestUpdateYamlKeyRejectsOversizeLine fails at the base (nil error, 8-
and 16-byte results) and passes with the guard. Adds a CHANGELOG
Fixed entry.

AC: AC-1
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin that a line just under the scanner limit still rewrites

The oversize-line refusal must fire only at the bufio.Scanner cliff. The
default scanner buffer holds bufio.MaxScanTokenSize (65,536) bytes, so a
line of 65,535 bytes still fits with its newline, and updateYamlKey has
to give a document with such a line the same rewrite a short document
gets: the key replaced in place when present, appended when absent.

Adds both rows to TestUpdateYamlKeyRejectsOversizeLine. They pass at the
base and with the AC-1 guard, which returns only on a scan error. Two
over-eager variants of the guard pass the existing yaml tests and the
AC-1 rows but fail these rows: capping the scanner buffer at
MaxScanTokenSize-1, and a pre-scan length check that refuses a line of
MaxScanTokenSize-1 bytes.

AC: AC-2
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin that a refused bd config set leaves config.yaml unchanged

setYamlConfigAtPath reads config.yaml, rewrites it through
updateYamlKey, and writes the result back. The AC-1 guard makes
updateYamlKey refuse a document with a line of bufio.MaxScanTokenSize
bytes or more, and setYamlConfigAtPath returns that error before its
os.WriteFile, so the file is never touched. No test covered that order.

TestSetYamlConfigAtPathOversizeLineLeavesFileUntouched writes a
config.yaml with a 65,538-byte comment line, sets no-db, and checks that
the error wraps bufio.ErrTooLong and that the file still holds all
65,563 of its bytes. At 5c9c02ab77 it fails both checks: a nil error,
and the file cut to the 11-byte `no-db: true`. It passes with the AC-1
guard and needs no code change of its own. A variant that writes the
file before returning the error fails the byte check (0 bytes left), and
one that swallows the error fails the error check.

AC: AC-3
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(changelog): say set-many keeps keys it wrote before a refused one

The entry said `bd config set` and `set-many` both fail and leave
config.yaml as it was. That holds for `set`, and for a `set-many` whose
first config.yaml key is the refused one. But `set-many` writes its
config.yaml pairs one at a time, in argument order, and returns on the
first error with no rollback, so a key listed before the refused
undotted key (a dotted `sync.remote`, say) is already on disk when the
command exits 1.

Measured at 0d3aa63761 with a 70,000-byte comment line in config.yaml:
`bd config set-many sync.remote=git@example.com:o/r.git actor=bob`
exits 1 and adds the `sync:` block (70,039 -> 70,083 bytes);
`set-many actor=bob sync.remote=...` and `bd config set actor bob` exit
1 and leave the file byte-identical.

de330c3e7f's message makes the same claim; this commit does not rewrite
it.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-19 09:15:48 +00:00
gascity-eddie[bot]
395e7beec0
fix(config): stop bd config set dropping the file's trailing newline (#7669)
* fix(config): keep config.yaml's trailing newline on bd config set

updateYamlKey's flat-key path, which `bd config set` takes for every
single-segment key, read the document with bufio.Scanner and joined the
lines back with "\n". ScanLines drops the final terminator, so the result
was exactly one newline short of any terminated config.yaml:
"issue_prefix: vp\ndolt.mode: server\n" came back without its final
newline, a spurious no-newline-at-end-of-file change in a git-tracked file.

Re-attach the input's own trailing run at the return, the tail
commentOutYamlKey got for the unset side in #6749. One statement covers
both the in-place and the appended branch, and a document without a
trailing newline still does not acquire one. The TODO that recorded the
defect is removed in the next commit.

The test file's header now covers both verbs; the set test follows the
unset ones.

AC: AC-1
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* chore(config): drop the TODO that deferred the bd config set newline fix

The TODO above updateYamlKey's return recorded that `bd config set` was
knowingly left one newline short, as a scope call in #6749. The previous
commit fixed that, so the TODO now describes a defect that no longer exists.

Check: searching internal/config/yaml_config.go for the phrase
"`bd config set` is knowingly left unfixed" finds one match at the base
and none after this commit.

AC: AC-11
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin a set's blank-line trailing run, not just a final newline

A config.yaml that ends in one or two blank lines came back from
`bd config set` one newline short, so the output still ended in "\n".
Compare the exact trailing run, as the unset table does, so that loss
stays visible. Both rows fail on the base code (run "\n" vs "\n\n", and
"\n\n" vs "\n\n\n") and pass with the tail fix.

AC: AC-2
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin the appended key's terminator on bd config set

When the key is absent, updateYamlKey appends it as the new last line,
and the old join left that line unterminated: "x\n" came back
"x\n\nnewkey: \"v\"". Pin the exact output for "x\n", where the separator
blank line is kept and the newline follows the new key, plus a table row
for the trailing run. Both fail on the base code and pass with the tail
fix.

AC: AC-3
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin a set appended over a blank-line tail

For "x\n\n" without the key, the existing blank line already separates,
so updateYamlKey adds no second one and the document's own two-newline
run now follows the new key: "x\n\nnewkey: \"v\"\n\n". Restoring a single
newline would leave this exact shape one short again, so pin it exactly,
plus the matching table row. Both fail on the base code, which returned
"x\n\nnewkey: \"v\"", and pass with the tail fix.

AC: AC-4
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin that a set does not invent a trailing newline

The repair re-attaches the document's own trailing run instead of appending
a newline, so a config.yaml that ends without one still ends without one
after `bd config set`. Pin that for both branches: a present key rewritten
in place, and an absent key appended as the new last line.

This passes on the base as well; it is a pin on "preserve, not append".
The always-append mutant (`return out + "\n", nil`) turns both inputs red.

AC: AC-5
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin a set's output on a CRLF config.yaml

A set of a single-segment key has always rewritten a CRLF document to LF,
because the scan drops each line's "\r". What the old join added on top was
a missing terminator: "actor: a\r\nother: x\r\n" came back
"actor: \"b\"\nother: x". With the restored tail it comes back
"actor: \"b\"\nother: x\n", which is also how an unset of the same document
ends.

Pin the exact output, and check the trailing run against commentOutYamlKey
on the same input, in a subtest named after unset's CRLF subtest. It fails
on the base (exact and parity) and passes with the tail from AC-1.

AC: AC-6
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(config): keep a CRLF config.yaml's blank-line tail on bd config set

The restored tail measures the trailing run on the raw content, where
strings.TrimRight(content, "\n") stops at the last "\r". On a CRLF file the
re-attached run was therefore always a bare "\n", however many blank lines
followed: "actor: a\r\nother: x\r\n\r\n\r\n" came back "…x\n", one newline
shorter than even the base's "…x\n\n".

Normalize CRLF the way commentOutYamlKey does before the scan, so the run is
the LF run and a set ends the file exactly as an unset of the same document
does. The scan output is unchanged, since ScanLines already drops each
line's "\r". The line sits below the nested dispatch, not at the top: some
updateNestedYamlKey branches carry "\r" through on untouched lines, and
this fix leaves that path alone.

Also correct the comment on unset's CRLF subtest, which still blamed a
leftover "\r" for the bare-"\n" tail; commentOutYamlKey normalizes up front,
so that mechanism no longer exists and the new set cases contradict it.

AC: AC-7
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin the trailing newline through SetYamlConfig

setYamlConfigAtPath writes updateYamlKey's output verbatim, so the dropped
newline reached the git-tracked config.yaml. Add the writer-level twin of
TestUnsetThroughTheFileKeepsTheTrailingNewline: a .beads/config.yaml ending
in a blank line, BEADS_DIR pointed at it, SetYamlConfig("actor", "bob"),
and the re-read file must keep its "\n\n" run. It first checks that the
value landed, so an untouched file cannot pass.

On the base the run comes back "\n"; with the tail from AC-1 it is "\n\n".

AC: AC-8
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(config): pin that a set of a dotted key bypasses the CRLF normalization

The CRLF normalization in updateYamlKey sits below the nested dispatch on
purpose. replaceFlatKeyLine and appendNestedYamlKey keep "\r" on the lines
they leave alone, so hoisting the line above the dispatch would silently
change what a dotted-key set writes:
"issue_prefix: vp\r\ndolt.mode: server\r\n" comes back
"issue_prefix: vp\r\ndolt.mode: \"embedded\"\n" today and would lose the
first "\r".

Compare updateYamlKey with updateNestedYamlKey on four CRLF documents for
dolt.mode: a literal dotted key, a nested block, an absent key, and a
comment-only document. This passes on the base as well; it pins the
boundary, not the nested writer's outputs. The hoisted mutant turns the
literal-dotted-key and comment-only rows red; the other two are re-marshalled
to LF either way.

AC: AC-9
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(config): append an absent key after a blank-line tail as one line

updateYamlKey's flat path appends an absent key after the scan's empty
tokens for the file's trailing blank lines, so those blank lines already
sit ahead of the new line; the tail rule then re-attached content's whole
run after it as well. A set of a new single-segment key on a file ending
in blank lines therefore repeated them at EOF: "x\n\n" came back
"x\n\nnewkey: \"v\"\n\n", "x\n\n\n" came back
"x\n\n\nnewkey: \"v\"\n\n\n", and CRLF "x\r\n\r\n" came back
"x\n\nnewkey: \"v\"\n\n" -- the diff noise this PR exists to remove.

The appended branch now re-attaches only the one terminator
bufio.Scanner dropped, and only when content had it; the in-place branch
keeps the TrimRight + content's-run rule it shares with
commentOutYamlKey. The appended-key subtest now wants
"x\n\nnewkey: \"v\"\n", adds "x\n\n\n" and CRLF rows, and drops the
comment claiming a single terminator would leave it one newline short.
The suffix-run table drops its blank-tail append row, whose expectation
assumed the file's run follows the key, and
TestSetOfAnAbsentKeyThroughTheFileAddsOneLine pins the shape through
SetYamlConfig. The doc comment states the nested boundary and the tail
rule.

Not changed here: the unchecked scanner.Err() is #7670's fix (its guard
lands right after this scan loop, and ./internal/config/ is green on the
union of both PRs). The stale bufio.Scanner narration in
commentOutYamlKey's comment stays a follow-up, since #7613 edits that
function.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-19 08:49:00 +00:00
Julian Knutsen
eb5ab48950
test(setup): pin the Claude agent artifacts bd init and bd setup claude write (#7678)
* test(setup): pin the Claude agent artifacts bd init and bd setup claude write

Before-parity pin P11. Nothing pinned the agent artifacts byte for byte, so
a change to them -- a new hook event in .claude/settings.json, new text in
the minimal CLAUDE.md/AGENTS.md template and with it a new marker hash that
marks every installed section stale -- could land without anyone seeing
the user-visible diff. These goldens make any such change a reviewed diff.

- internal/templates/agents: TestManagedSectionGolden pins each managed
  section (minimal and full profile, with and without a Dolt remote)
  markers included; TestManagedSectionHashes pins their marker hashes as
  hand-maintained literals; NoPush is pinned to render the no-remote bytes.
- cmd/bd/setup: TestClaudeSetupArtifactsGolden records transcripts of
  bd setup claude, --check and --remove (project, --stealth, --global, a
  user's own settings and CLAUDE.md, a stale section, a legacy
  settings.local.json install, a CLAUDE.md that imports AGENTS.md): each
  step's result and output, and the settings and instruction files it
  creates, changes or deletes. TestClaudePluginManifestHookEvents pins the
  hook events and commands the beads Claude plugin declares.
- cmd/bd: TestEmbeddedInitAgentArtifactsGolden runs a real bd init (default
  and --agents-profile full) and then bd setup claude --check, --remove and
  a reinstall through the binary, with exit codes.

A managed section in a transcript is shown by its marker line and a
pointer to the template golden it is checked to match byte for byte, so a
template change shows up once in full and as a hash bump wherever it lands.
Regenerate with BD_UPDATE_GOLDEN=1.

CI companions: the new test files and testdata in BUILD.bazel (bd_test
gains its internal/templates/agents dep), and the new TestEmbedded* in the
100-shard embedded cmd manifest (incremental --write).

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(setup): drop BD_OTEL_* and split HOME in the agent artifact golden

TestEmbeddedInitAgentArtifactsGolden pins bd's stderr byte for byte, but
bdEnv passes the caller's BD_OTEL_* variables through, and bd prints a
deprecation warning on stderr for a legacy one (BD_OTEL_METRICS_URL,
BD_OTEL_LOGS_URL, BD_OTEL_STDOUT=true) before it checks whether
telemetry is on. On a host that exports them (agent hosts do; CI does
not), both scenarios failed at the first stderr block with no code
drift.

The test now runs bd init and every bd setup claude step with
agentArtifactEnv: bdEnv(dir) without BD_OTEL_* and OTEL_*. It filters
the env slice instead of calling clearTelemetryEnv, which uses t.Setenv
and so cannot run in these parallel tests.

The same helper gives bd a HOME of its own instead of bdEnv's HOME ==
project. With the two equal, a bd that wrote the hook and CLAUDE.md into
the home directory instead of the project passed every pin. The
transcript now also records $HOME/.claude/settings.json and
$HOME/CLAUDE.md, and output paths under HOME read $HOME. bd init runs
the steps bdInit does (git repo, seeded schema) with that env.

The test doc now names what it does not pin: a proxied-server bd init
(no Codex section, HasRemote from its own remote URL), a bd init with a
remote, and the Cursor files. It also says a change to AGENTS.md's
wrapper text or Codex section can fail only here.

The goldens are unchanged.

Validation (env -i, go1.26.7, GOFLAGS=-mod=readonly, GOPROXY=off,
BEADS_TEST_EMBEDDED_DOLT=1, -tags gms_pure_go):
- the golden passes in a clean env; with BD_OTEL_METRICS_URL=x
  BD_OTEL_LOGS_URL=y; with every legacy selector plus BD_OTEL_ENABLED
  and OTEL_EXPORTER_OTLP_ENDPOINT; and under -race;
- the previous test with BD_OTEL_METRICS_URL=x BD_OTEL_LOGS_URL=y fails
  at golden line 132 on the deprecation warning;
- swapping homeDir and projectDir in setup's defaultClaudeEnv passes the
  previous test and fails this one at golden line 5 (the transcript
  shows created $HOME/.claude/settings.json).

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-19 05:19:28 +00:00
Julian Knutsen
e24940e0c1
fix(workspacegate): strict --readonly never creates workspace gate files (#7672)
* fix(workspacegate): strict --readonly never creates workspace gate files

Strict --readonly is mutation-free by policy (effectiveRootStorePolicy),
but the command chokepoint took its SHARED workspace gates with
O_CREATE|O_RDWR. On a workspace no bd command had gated yet, a strict
readonly command therefore left <project>/.beads.gate.lock and the
physical-root gate file (.beads/embeddeddolt.gate.lock or
.beads/dolt.gate.lock) behind.

Add workspacegate.Options.ReadOnly. A Shared acquisition with it opens
the gate file O_RDONLY (the idiom ExclusiveHolder already uses) and never
creates it. A gate file that does not exist means nothing has ever been
gated there, so Acquire succeeds with an empty handle whose Release is a
no-op. An existing gate file is held SHARED as before, so maintenance
still sees the command and a live exclusive holder still makes it wait
and then refuses it. The option is ignored for Exclusive: maintenance has
to be able to create the file it locks.

ReadOnly composes with the fair, bounded waiting from #7225 without
changes to it: the writer-fairness probe (ExclusiveQueued) already opens
the intent file read-only and treats a missing one as nothing queued, so
a waiting ReadOnly acquisition creates no .intent or .info file either.
readonly_wait_test.go pins that seam.

The chokepoint passes ReadOnly: readonlyMode in the SHARED branch. Every
other command keeps creating the gate files, which is what makes it
visible to maintenance. Residual, documented at the chokepoint: on a
workspace that has never been gated, a strict readonly command is
invisible to a maintenance operation that starts while it runs.

Ported from the versioned-beads integration fork (donnabox):
0c86e2b205 (red tests), bac6ff95dc (fix), b7a4f96a5a (fair-wait tests),
2af5a664dc (BUILD.bazel registration), and the #7225 conflict resolution
in 9d128f8c51. The fork's `--readonly backup restore` refusal
(6ac79d6207) is a separate change and is not included.

Co-authored-by: Jim Wordelman <jim@wordelman.name>
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(cli): release the previous command's workspace gate before acquiring

acquireCommandWorkspaceGates stored its new handle over whatever
workspaceGateHandle already held. cobra skips PersistentPostRunE when
RunE fails, so in a process that runs more than one command (tests,
embedders) a command could start with the previous command's handle
still set. Two things then went wrong:

- on the paths that take no new gate (no .beads on disk, a failed or
  fail-open acquisition) the stale handle stayed set, so the previous
  workspace stayed held for the whole command;
- on a successful acquisition the stale handle was overwritten without
  being released, so its flock stayed held on the previous workspace
  until the garbage collector closed the file, and a maintenance
  operation there was turned away by a command that had finished.

Release any held handle first. In a single-command bd process the handle
is always nil at that point, so this is a no-op there.

Ported from the versioned-beads integration fork (donnabox): 3a25b42ed1
(cherry-pick of ad87442c92). The test
TestAcquireCommandWorkspaceGatesReleasesThePreviousCommandsHold is new
here and covers the successful-acquisition case the fork's tests do not.

Co-authored-by: Donna Box <dbox@boxenllc.com>
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(cli): close a stale store before the leading gate release

Review fixups for #7672, apply-fixes iteration 1.

The leading releaseWorkspaceGates() that the second commit added to the
chokepoint was the only production release without a
closeStoreBeforeGateRelease() first. When an in-process command's RunE
fails, Cobra skips PersistentPostRunE. The next command in the same
process then dropped that command's gates while its store and provider
were still open, and later overwrote and leaked them. Close them first,
as every other release path does. The current command assigns
store/uowProvider only after the chokepoint, so anything set there
belongs to the previous command. Only multi-command processes (tests,
embedders) are affected; production runs one command per process.

A strict --readonly hold on a never-gated workspace locks nothing. The
chokepoint still exported this PID in BEADS_GATE_SHARED_HOLDER_PID, so bd
children skipped the writer-fairness queue behind a parent that blocks
no exclusive acquirer. Add Handle.Empty and MultiHandle.Empty, and
advertise only when at least one gate is actually locked.

Tests:
- A regression test runs the chokepoint over a stub store and provider.
  It pins that both close while the previous command's gate is still
  held.
- A marker test forces advertiseSharedHoldEnabled on. It pins that a
  hold of nothing is not advertised, and that the hold is advertised
  once the gate files exist.
- The strict-readonly chokepoint subtests now require a handle that
  holds nothing. SHARED fail-open leaves no handle, so these subtests
  now fail the review's W2 mutant, which passed them before.
- A package test covers Empty.

The posture comment now says the gate follows the strict flag only.
Allowlisted reads and --dry-run/--inspect previews still create the gate
files.

Not changed: `bd --readonly backup restore` still creates gate files on
its exclusive acquisition. This PR covers shared gates only. Sibling #7673
refuses a strict-readonly restore before it takes the exclusive gate.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Jim Wordelman <jim@wordelman.name>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Donna Box <dbox@boxenllc.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-19 04:31:00 +00:00
Julian Knutsen
05fffb1e11
test(storage): versioned-history flag-off differential and statement pins (#6132, #6135) (#7681)
* test(domain/db): pin the statements the history-off upsert Insert issues

The participation_generation write fence (#6943) gave domain/db Insert's
upsert path (CreateOnly unset: the quick, todo, gate and batch proxied
servers' create) an existence probe ahead of the upsert, skipped while
versioned history is off. Pin that skip with a sqlmock statement-sequence
test: with history off, the upsert into issues and into wisps issues
exactly the three statements the pre-stack code (4a67d0d028) issues --
the upsert, the created event's same-content scan and its insert -- and
nothing else.

The mock is reached through internal/storage/sqlcount, and the pin also
requires exactly three statements to reach the driver. sqlmock alone
answers an unscripted statement with an error and fails only if the
caller returns it, so a best-effort read whose error Insert dropped would
pass its checks; the count catches it. The history-on run of the same
script is the control: the probe (and the mint after it) must make the
pin refuse it, which shows the script is exhaustive and that Insert reads
the transaction's scope.

Passes on 4a67d0d028 and on main; main issues no extra read with history
off. Before-parity pin P2 (Insert half) from the fork gap analysis, for
gastownhall/beads#6132 / #6135.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(issueops): pin the statements the history-off defer-wake sweep issues

Port the fork's sqlmock pin of WakeExpiredDefersInTx with versioned
history off (donnabox/codex/preview3-source-catchup-46aa 11c74cb575):
a one-row sweep issues exactly the expired-row SELECT, the UPDATE, the
event's same-content SELECT and INSERT, and the empty wisps pass, so any
read added before or after waking a row while history is off fails the
pin. It passes on main unchanged and guards the fork's wake-skip change,
which checks each expired row against the mint before waking it and must
keep that check free with history off.

Adapted for main: the doc no longer describes the fork's pre-wake check
as existing; a history-on run of the same script is added as the control
(the mint's read must make the pin refuse it), so the pin cannot pass
vacuously; and the mock is reached through internal/storage/sqlcount, so
the pin also requires exactly five statements to reach the driver. That
catches a read whose error the sweep swallowed, which sqlmock alone does
not see. Passes on 4a67d0d028 and on main. Before-parity pin P2 (wake
half), for gastownhall/beads#6132 / #6135.

Co-authored-by: Jim Wordelman <jim@wordelman.name>
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(storage): versioned-history flag-off differential against pre-stack goldens

gastownhall/beads#6132 and #6135 require a differential test proving that
with versioned history off (the default) storage behaves byte-identically
to before the feature. Existing coverage stops short of that:
TestDualWriteContractFlagOff only counts version rows, the
participation-generation contracts check one column, and createbatchequiv
runs its batch with history on over a column subset.

Add internal/storage/historyoffequiv. Each leg (embedded Dolt, Dolt
server, unit of work) runs one scripted mutation set through its own
front doors on a store whose history was never turned on:

- on the issues plane, every write role: creates, batch create, apply
  batch, updates, the no-op update, reparent, dependency add/remove,
  comments, claim/heartbeat/release, lease reclaim, metadata CAS,
  close/reopen, batch close with a refused item, defer and wake, delete;
- the persistence moves in both directions;
- on the wisps plane, the roles again: create, update and the no-op
  update, dependency add/remove, comment, metadata CAS, close/reopen,
  apply batch, batch close, release, defer and wake (the sweep's wisps
  pass), delete, a forced close of a no-history wisp, and the refusals
  the lease roles' contracts give a wisp (Claim, Heartbeat; Reclaim finds
  nothing);
- each leg's own surface: the store legs' DoltStorage methods (import
  upsert, promotion, RunInTransaction, ...) and the uow leg's
  CreateOnly-unset domain upsert and importer.

Neither reclaim sweep finds an expired lease, since no role can age one.
The harness then dumps every column of every base table, plus the
scenario's Dolt commit messages and the final dolt_status, and compares
per-table SHA-256 digests with goldens recorded by running the same
harness against origin/main 4a67d0d028, the commit before the stack
(#6943, #6661, #6664).

The only tolerated differences are listed in schemaAdditions: the stack's
new columns (issues/wisps.participation_generation,
issue_versions.removed_restriction) must be NULL in every row and are
then left out of the digest; its new table (epoch_minted_addresses) must
be empty; its own migration bookkeeping rows (0070-0072, ignored/0028)
may each be there once. Recording and checking apply that rule the same
way, so a golden re-recorded from main is byte-identical to the
pre-stack one (verified on all three legs). The package doc says how the
goldens are kept: a later migration whose schema flag-off leaves
untouched adds its own schemaAdditions entry; a deliberate flag-off
change, a scenario change or an engine rendering change re-records them
and shows the dump diff. Wall-clock instants, wall-clock-derived ids and
random tokens are masked by explicit, documented rules found by diffing
two runs; instants the scenario supplies stay verbatim, so a write that
starts overwriting one still shows.

All three legs pass on main. Before-parity pin P1 from the fork gap
analysis; the new embedded test is placed in the 40-shard storage block.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Jim Wordelman <jim@wordelman.name>
2026-10-19 03:47:00 +00:00
Julian Knutsen
f6f122894f
test(serve): before-parity pins for the httpapi startup event and ordinary-workspace bd serve (P13, P14) (#7679)
* test(httpapi): pin the startup event's capabilities and database source

TestStartupLines and TestStartupNamesTheDatabaseSource check the
event=startup line by substring, and neither reads its capabilities
field. A server could print one capability list and advertise another
on GET /v0/beads/context, or a new kind of database source could
relabel the line, and both tests would stay green.

TestStartupEventPublishesTheAdvertisedCapabilitiesAndSource parses the
line back into its fields and pins, for the provider and roles sources
with and without the external-dependency policy: the exact field
sequence, db=provider|roles, and a capabilities value equal to the
handshake's list joined with commas. It is a before-parity pin: a
future alternative source on the same server must leave this line
unchanged for the Issue API's two sources.

Test-only; no behavior change.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(serve): pin ordinary-workspace bd serve under BD_GRAPH_MODE=dependency

bd reads no BD_GRAPH_MODE today. "dependency" names the format every
existing workspace already has, so any later reader of that variable
(a graph-mode serve dispatch, for example) has to keep serving an
ordinary workspace through runServe exactly as it does now. These
tests pin that before such a reader exists.

Each test runs bd serve twice in one workspace, once with the variable
unset and once with BD_GRAPH_MODE=dependency. It requires the
event=startup and event=limits lines to match byte for byte apart
from the ephemeral port, and the GET /v0/beads/context bodies to be
equal. Each run is also checked against what runServe builds: the
serveResolvedMode label, the db source, and a startup capabilities
field equal to the handshake's list. Two runs that were both routed
somewhere else therefore cannot pass by agreeing with each other.

- TestServeOrdinaryWorkspaceIgnoresGraphModeDependency: in-process
  through rootCmd (PersistentPreRunE included), on the store source of
  a registered-backend workspace. Runs in bd_test.
- TestServerModeServeIgnoresGraphModeDependency: a real bd on a
  `bd init --server` workspace, which uses the provider source. Gated
  on BEADS_TEST_PROXIED_SERVER like the other TestServerMode* tests,
  and placed in the proxied 34-shard manifest by
  gen_proxied_shard_manifest.py --write (incremental).

The new file is listed in cmd/bd/BUILD.bazel's bd_test srcs.
Test-only; no behavior change.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(serve): stop the in-process graph-mode pin hanging when serve fails

serveInProcessSnapshot's cleanup decided whether the server was still
running by looking at rootCancel. That holds for a clean run, which
clears rootCancel in PersistentPostRunE. A run whose RunE returns an
error skips PersistentPostRunE, so rootCancel stays set even though the
main flow has already read the run's only result from done. The
cleanup then called rootCancel and blocked on done for good, and the
failure surfaced as a test-binary timeout panic that takes the whole
bd_test shard with it instead of a FAIL for this test.

The cleanup now keys on whether done has been read. A serve that binds,
answers, and then returns an error (one way a future graph-serve
dispatch could regress) now fails in about 5s with "bd serve returned
<err>, want a clean shutdown". A failure before done is read (a bad
GET, say) still cancels the server and waits for it.

Test-only; no behavior change.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-19 03:26:45 +00:00
Jim Wordelman
5c9c02ab77
Phase 3 (R20): epoch-transition enforcement + minted-address CAS - gastownhall/beads#6136 (#6664)
* test(conformance): un-skip R16 CAS + R20 epoch tests with honest not-implemented stubs

Wires all 10 ExpectedRevisionFixture-gated Run* functions and both
EpochFixture-gated Run* functions to real (non-nil) hooks on all three
storage legs (dolt, embeddeddolt, uow), replacing the Phase 0 all-nil
fixtures that made these cases t.Skip.

Every wired hook is an unconditional "not implemented yet" stub rather
than a partial-real implementation. This is deliberate: every Run*
function under test calls its primary hook (CompareAndSetVersion, or
CurrentEpoch/MintUnderEpoch) first, before any other hook, so a stub
that errors unconditionally on that first call guarantees deterministic
RED with zero risk of an accidental pass -- a concern that's concrete,
not hypothetical, since at least one case
(RunExpectedRevisionAcceptsAWriteNamingNoVersion) would go green under
a partially-real read/unconditional-write strategy backed by Phase 1's
already-persisted current_revision/store_epoch columns.

Each case now fails for its documented reason: most via a direct
t.Fatalf quoting the stub's "not implemented" error, and two
(RefusalIsDistinguishableFromNotFound,
RefusalIsDistinguishableFromValidationFailure) via a clean want/got
mismatch against the sentinel errors they expect.

Explicitly out of scope and untouched: CrossRecordInvariantFixture
(R16.1) and the non-epoch RetentionFixture cases (R17) -- confirmed via
test run that TestRetentionContract still reports PASS with all 13
subtests SKIP on every leg.

TestEveryLegWiresEveryRoleContract (internal/storage/leg_contract_wiring_test.go)
is a source-level AST naming check independent of hook nil-ness; all 12
entrypoints were already called by every leg pre-change, so this lock
was satisfied before and remains satisfied after.

Sentinel error messages reference be-x5jqd.3 (R16 CAS implementation)
and be-x5jqd.4 (R20 epoch enforcement implementation), the two builder
children this un-skip unblocks.

Verified via go build ./..., go vet ./..., and go test -tags cgo across
all three legs: all 10 R16 + 2 R20 subtests FAIL for the documented
reason, TestRetentionContract unaffected.

Bead: be-x5jqd.1

* feat: green — Implement R20 epoch-transition enforcement (refs be-x5jqd.4)

* fix(epoch): honor R20-n's retained-mapping exception in StillServes/Resolve

StillServesInTx and ResolveEpochInTx only ever checked whether the
literal queried address's own row matched the current epoch. Per R20-n
(#5898 rev 9), a version that survives an epoch transition keeps its
prior-epoch address resolving once its id has been carried forward into
the current epoch (via CurrentAddressForInTx or a fresh
MintUnderEpochInTx) — that carried-forward row wasn't being consulted,
so a legitimately retained address would wrongly stop resolving.

Add mintedIDHasAddressAtEpochInTx and consult it as a fallback in both
functions when the queried address's own row is stale. Addresses whose
id was never carried forward are unaffected (row.mintedID has no row
at the current epoch, same false/GoneReorganization answer as before).

Round-2 response to be-rskwc's causal_disproof finding: the merge-base
commit FAILed honestly on this defect, this branch's tdd_green SKIPped
instead (attribution_evidence). This fix is real and verified via a new
sqlmock unit test (TestStillServesInTxHonorsARetainedMapping et al.),
but does not by itself flip
TestEpochContract/EpochBumpVoidsOnlyAddressesOfVersionsNoLongerServed
from SKIP to PASS: that test's own construction mints two records
identically and never carries either forward before comparing
StillServes on both, so no backend implementation can make them differ
(escalation follows on a needs-mayor bead).

refs be-x5jqd.4, be-rskwc round 2

* test(feat): red — harden EpochBumpVoidsOnlyAddressesOfVersionsNoLongerServed (refs be-9fiaw)

Per mayor ruling on be-9fiaw: fix the test rather than waive it. The
fixture already lets record-a survive an epoch bump by minting it again
under the new epoch; record-b, never re-minted, does not survive. That
asymmetry was previously only discovered after the fact (skip if
StillServes agreed for both addresses), which meant the case could pass
by observing either polarity. Assert the concrete expected values
instead: servesA must be true, servesB must be false. This makes the
case deterministically exercise both halves of R20-n, and closes a
latent gap where a backend with the polarity reversed could have
silently passed via the old discover-then-swap logic.

* fix(schema): renumber epoch_minted_addresses to 0070, linearizing our own slot collision (refs be-hs42e.7)

R20's epoch_minted_addresses and R7.1's removed_restriction both claimed
migration slot 0069 on parallel sibling branches. Neither contained the
other, so nothing in either branch's own CI could see the union: each was
green alone, and the second to merge would panic checkNoDuplicateVersions at
store open -- before any command's RunE, so every bd invocation bricks rather
than one package reddening. Same hazard bee reproduced for #6358 vs #6008.

This linearizes them instead of leaving it to merge order. R7.1 keeps 0069:
it is published beneath #6661 and cannot move without rewriting a branch a
reviewer is reading. R20 stops being a sibling, rebases onto R7.1, and takes
the next contiguous slot.

Contiguity is why the renumber could not be done pre-emptively in either
branch alone: TestAllMigrationsSQLAppliesThroughDoltCLIAndRecordsLatestVersion
asserts COUNT(*) == LatestVersion(), so a jump to a far slot fails until the
gap is filled. The order had to be settled first.

- migrations/0069_add_epoch_minted_addresses.{up,down}.sql -> 0070_*
- migration_0069_add_epoch_minted_addresses_test.go -> migration_0070_*, with
  TestLatestVersionIncludesMigration0069 -> ...0070 (want 70)
- 0069_add_removed_restriction_test.go retires its own LatestVersion pin in
  favour of 0070's, keeping the one-pin-at-a-time chain 0068 -> 0069 -> 0070

cli_migrations.go needs no change: its only override is R7.1's guarded
PREPARE, which keeps 0069. The epoch migration is a plain unguarded
CREATE TABLE IF NOT EXISTS and needs none.

Verified: TestEmbeddedMigrationSourcesHaveNoDuplicateVersions,
TestLatestVersionIncludesMigration0070,
TestAllMigrationsSQLAppliesThroughDoltCLIAndRecordsLatestVersion (applies the
full 0001..0070 chain through the dolt CLI),
TestBundleMigrationsWithPreparedALTERAreOverriddenOrJustified and
TestAllMigrationsSQLUsesDirectDDLForKnownCLIIncompatibilities all pass.
go build ./... and go vet ./... clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* test(feat): red — Fix PR #6664: restore R16 honest-skip (stale be-x5jqd.3 premise) + address bee's 2 blockers on epoch_cas.go (refs be-ah0jf)

* feat: green — Fix PR #6664: restore R16 honest-skip (stale be-x5jqd.3 premise) + address bee's 2 blockers on epoch_cas.go (refs be-ah0jf)

* test(feat): red — Review: Fix PR #6664: restore R16 honest-skip (stale be-x5jqd.3 premise) + address bee's 2 blockers on epoch_cas.go (refs be-yj55d)

* feat: green — Review: Fix PR #6664: restore R16 honest-skip (stale be-x5jqd.3 premise) + address bee's 2 blockers on epoch_cas.go (refs be-yj55d)

* docs(epoch): cite epoch_minted_addresses as migration 0071

Upstream's 0069_widen_issue_versions_datetime_precision (#6675) took slot
0069 first, so removed_restriction moved to 0070 and epoch_minted_addresses
to 0071. The three epoch_cas.go headers still named 0070.

Rebase note (maintainer): the migration is now 0072, per the landing-order
ruling (#6943's participation_generation is 0070 and #6661's
removed_restriction is 0071); the commit "chore(schema): renumber
epoch_minted_addresses to 0072 after #6661's 0071" updates these three
headers.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(feat): red — move #6664's Dolt server leg to post-transaction publish for store_epoch and epoch_minted_addresses (latent lost-update pattern)

Tests only. On the commit before this one they fail for these reasons:

- the server leg publishes store_epoch / epoch_minted_addresses inside the open
  SQL transaction (DOLT_ADD before COMMIT) for BumpEpoch, MintUnderEpoch and
  CurrentAddressFor: the lost-update hazard documented on doltAddAndCommitInTx
- MintUnderEpoch and CurrentAddressFor stage only epoch_minted_addresses, so the
  first mint leaves store_epoch modified in the working set, on both legs
  (#6664 Major 2, reproduced with Status() on a real server and embedded)
- BumpEpoch must surface a Dolt commit that fails after its SQL commit, while
  the mint paths keep their rows
- ensureStoreEpochRow seeds with INSERT IGNORE and reads the row back

TestMintOverlappingABumpCommitsAnAddressThatIsAlreadyGone characterizes current
behaviour and passes now: a mint overlapping a bump commits an address the bump
already voided. Its doc comment records why that is left as it is.

* feat: green — move #6664's Dolt server leg to post-transaction publish for store_epoch and epoch_minted_addresses (latent lost-update pattern)

Server leg (internal/storage/dolt/epoch_cas.go): BumpEpoch, MintUnderEpoch and
CurrentAddressFor no longer stage and Dolt-commit inside the open SQL
transaction. The two mint paths publish through runIssueOperationTxWithMessage,
which commits the SQL transaction first. BumpEpoch commits it, then publishes
with doltAddAndCommitPostTx and RETURNS a failed Dolt commit: a bump that is
committed but unpublished must not read as success.

Both legs: a mint stages store_epoch with epoch_minted_addresses, because the
first mint inserts the singleton row (#6664 Major 2). The set lives in
issueops.EpochMintDirtyTables so the legs cannot drift from what the body writes.

issueops: ensureStoreEpochRow seeds with INSERT IGNORE and reads the row back,
as version_history.go does for the same row.

Not changed, on purpose: a mint overlapping a concurrent bump can still commit an
address the bump already voided. Dolt has no lock for it and the only conflicting
write would overwrite bumped_at/bumped_reason; see the doc comment on
TestMintOverlappingABumpCommitsAnAddressThatIsAlreadyGone. One comment in that
test file is reworded; no test logic changed since the red commit.

* chore(bazel): bazel-sync (gazelle) for the two new epoch test files

Only the two go_test srcs entries for the test files this branch adds:
epoch_cas_publish_test.go (dolt) and epoch_cas_status_test.go (embeddeddolt).
gazelle wanted nothing else for them (no deps change).

* test(conformance): add the optional EpochFixture.LoseVersion capability

Record survival explicitly instead of inferring it (gastownhall/beads#6664,
review 5360880888, Major 1). Once an address stays served unless its own row
says the store lost it, the only way for the contract to produce a Gone
address is an explicit loss, so the contract needs a way to say "the store
lost this Version".

LoseVersion is additive and optional: no existing field changes signature,
and a nil LoseVersion makes the cases that need one skip with a named reason.
It is its own commit so that it can be judged, or replaced, on its own: the
same semantics could ride on BumpEpoch taking a lost set, and only the
plumbing would differ. Nothing uses it yet.

* chore(epoch): add signature-only stubs for the survival rework

The tests that follow exercise BumpEpochCarryingInTx, LoseVersionInTx,
ErrEpochAddressNotServed and the BumpEpochCarrying and LoseVersion store
methods (gastownhall/beads#6664, review 5360880888). Stubs give those names
so the tree builds and the tests fail for their real, behavioural reasons
instead of failing to compile.

No logic: ErrEpochAddressNotServed is a sentinel, BumpEpochCarrying advances
the counter exactly as BumpEpoch does and carries nothing, and LoseVersion
returns "not implemented". The next commit adds the tests; the one after
replaces these bodies.

* test(feat): red — record survival explicitly in epoch_minted_addresses, make CurrentAddressFor read-only, carry on a token-scheme change

Tests for the survival rework of gastownhall/beads#6664 (review 5360880888:
Major 1, Minors 3 and 4). Every test is written before any production change.

- Contract: the earlier mint-driven R20-n case is restated as three named
  phases (a survivor and a lost Version, loss is terminal, survival needs no
  upkeep), plus three new cases for a token-scheme change (carry the live
  lineages and only those, a carried address keeps resolving across later
  bumps, a lost address stays Gone across a scheme change). All three legs
  register them and wire LoseVersion and the carrying bump.
- Shared body: unit tests for the row-based resolver (exactly the address's
  row and the epoch, fail closed on a row from a future epoch), a read-only
  CurrentAddressFor with typed errors, LoseVersion, a bump that writes only
  store_epoch, a mint that never hands back a lost address, and the
  per-database counter (Minor 3). Minor 4 gets a guard that no address in the
  test file is written by hand.
- Real engine: the carry on the embedded leg (one carry row per live id, the
  lineage root, the collision rule, a row from a future epoch, a rewound
  counter, database-wide, lost lineages), the publication sets of every
  writer, and on the server leg the overlap of a mint with a bump (now a Live
  survivor) and a deterministic mint-after-loss interleaving that must not
  revive the lost address.
- Migration 0071 byte pin and a real-dolt shape check for carried_from,
  gone_at_epoch, the carried_from index and the gone-not-before-mint CHECK.

These fail on the current code, and some do not build yet because
BumpEpochCarryingInTx, LoseVersionInTx, ErrEpochAddressNotServed and the store
methods they exercise do not exist.

* feat: green — record survival explicitly in epoch_minted_addresses, make CurrentAddressFor read-only, carry on a token-scheme change

The survival rework of gastownhall/beads#6664 (review 5360880888: Major 1,
Minors 3 and 4). An address's status is now a function of its own row and the
current epoch, never inferred from other rows.

- Migration 0071 (edited in place; no new slot, no ALTER): the table gains
  carried_from, gone_at_epoch, an index on carried_from and a CHECK that a loss
  epoch is not before the mint epoch.
- Shared body: Live unless gone_at_epoch is set, and a row from an epoch that
  has not happened yet fails closed; the chain query is deleted. CurrentAddressFor
  only reads and returns typed errors. LoseVersion marks a whole lineage lost,
  once, guarded by gone_at_epoch IS NULL. BumpEpochCarrying (token-scheme change)
  writes one carry row per live lineage naming the lineage root, resolves a
  shared-id collision by carrying the newest and losing the rest, and fails if a
  row already sits at the new epoch. A mint only inserts its own root row and
  never hands back an address that is not served. A restore or reinit bump
  still writes only store_epoch.
- Both Dolt legs: BumpEpochCarrying publishes store_epoch and
  epoch_minted_addresses, LoseVersion publishes epoch_minted_addresses and
  returns a failed publication, and CurrentAddressFor runs on the read path and
  publishes nothing. The per-database epoch counter (Minor 3) and the
  requirement that a caller of the carrying bump hold the workspace gate
  exclusively are documented where they apply.

One change to tests written in the previous commit: the three real-server
publication probes that only need a clean working set now open a database of
their own, as the overlap tests do, instead of queueing on the shared test
database's schema-init lock. Its five-second acquire timeout ran out under load
and failed two of them; no assertion changed.

* test(dolt): run the epoch contract on a database of its own

TestEpochContract set up through setupTestStore, which puts every parallel
test on the shared test database and queues their schema initialization on one
lock with a five-second acquire timeout. Under host load a setup that waits
behind another parallel test's runs out of it and fails at its first line with
"schema migration lock unavailable: timeout". That failed the unchanged
first-mint test twice in the dolt lane once this change had added tests to it,
and the contract's own setup is exposed to the same timeout.

The contract now opens a database of its own, the way the overlap tests and the
publication probes already do, so this change adds no contention on the shared
database's lock. No assertion changed.

* docs: replace internal tracker ids in the epoch comments and test names with plain descriptions

Comment and text edits only; no behavior changes and no line is added or
removed (21 lines are replaced one for one). Each replaced line named an
internal tracker id in a header, a test comment, a failure message or the
migration header; it now says what it meant in plain words or cites the
upstream issue or pull request. The EpochFixture API lines and the already
published commit messages are left as they were.

* ci: list this PR's embedded storage tests in the embedded shard manifest

#7156 made the embedded-Dolt tier's shard manifests exact-once (50 cmd
shards, 15 storage shards), so every embedded test must be listed. This PR
adds eleven embedded storage tests and no embedded cmd tests; this commit is
exactly what the generators write for them:

  python3 scripts/ci/gen_embedded_cmd_shard_manifest.py 50 --weights=duration --write
  python3 scripts/ci/gen_embedded_storage_shard_manifest.py 15 --weights=duration --write

Incremental: 11 lines added, none moved; the cmd manifest is unchanged.

Rebase note (maintainer): regenerated on current main, whose embedded storage
manifest uses a 40-shard block; the generator's incremental output placed
these tests at shards 7, 9, 11, 20, 21, 28, 31, 32, 33, 34 and 36 of 40, one
test per shard (gen_embedded_storage_shard_manifest.py 40 --weights=duration
--write; --check passes).

* chore(bazel): bazel-sync for the epoch sources and migration 0071

The BUILD files did not list epoch_cas.go in the dolt, embeddeddolt and
issueops libraries, issueops' epoch_cas_test.go, or migration 0071 and its
test, so Bazel could not compile those packages' tests. This is exactly the
patch make bazel-sync writes (gazelle plus go_srcs.py; the PR's BUILD-sync
CI step printed it): 4 files, 6 lines, nothing else.

Rebase note (maintainer): the migration is now 0072, per the landing-order
ruling (#6943's participation_generation is 0070 and #6661's
removed_restriction is 0071); the commit "chore(schema): renumber
epoch_minted_addresses to 0072 after #6661's 0071" renames these BUILD
entries. On the rebased base, this commit's internal/storage/schema/BUILD.bazel
hunk is merged with the base's 0070 and 0071 entries.

* chore(schema): renumber epoch_minted_addresses to 0072 after #6661's 0071

Migrations are numbered by landing order. #6943 (participation_generation,
the versioned-history write fence) owns main-plane 0070 and #6661
(removed_restriction) owns 0071, so this PR's epoch_minted_addresses
migration moves to the next free slot, 0072. The rename is done here, as one
maintainer commit, rather than inside the replayed commits, so the history
shows exactly what changed relative to the PR as reviewed.

- migrations/0071_add_epoch_minted_addresses.{up,down}.sql become 0072_*.
  Neither header names its own number, so the SQL and its comments are
  unchanged. Neither file exists on main, #6943 or #6661, so this is not an
  edit to a shipped migration.
- cli_migrations.go needs nothing: the migration is a plain CREATE TABLE IF
  NOT EXISTS with no CLI-bundle override.
- migration_0071_add_epoch_minted_addresses_test.go becomes
  migration_0072_add_epoch_minted_addresses_test.go; its consts, tests and
  messages say 0072, it owns the one LatestVersion pin (72), and its slot
  history now records the move past #6943's 0070.
  TestLatestVersionIncludesMigration0071 in #6661's removed_restriction test
  was already retired to a pointer comment by the rebase's resolution of
  "renumber epoch_minted_addresses to 0070"; that comment now names
  TestLatestVersionIncludesMigration0072 and its file.
- BUILD.bazel embedsrcs/srcs, and the three epoch_cas.go headers (dolt,
  embeddeddolt, issueops) that cite the migration number.

The intermediate commits of the rebase are not expected to build, for two
reasons:

- a duplicate migration version: the PR's epoch_minted_addresses migration
  shares 0069 with main's 0069_widen_issue_versions_datetime_precision until
  the PR's own "renumber epoch_minted_addresses to 0070", and from there
  shares 0071 with #6661's removed_restriction until this commit;
- a helper-name collision: from the PR's "feat: green — Fix PR #6664: restore
  R16 honest-skip ... (refs be-ah0jf)" on, internal/storage/issueops declares
  readStoreEpochInTx twice, once in #6943's version_history.go and once in
  this PR's epoch_cas.go, until "fix(epoch): rename #6664's read-only
  store_epoch read past #6943's readStoreEpochInTx" later in this stack.

This commit ends the duplicate version, but it still does not compile on its
own because of the collision; the stack builds from the rename commit on.

Checked at the tip of the stack, where the rename has landed: go test
./internal/storage/schema/... with the dolt CLI lane
(including TestAllMigrationsSQLAppliesThroughDoltCLIAndRecordsLatestVersion
and TestMigration0072RecordsSurvivalInTheAddressTableThroughDoltCLI), and
scripts/check-migration-hygiene.sh with BASE_SHA=origin/main and
BASE_SHA=vh/6661-rebase.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(epoch): waive #6664's three new epoch contracts on main's http leg

main gained an HTTP storage leg (#7247) after this PR's base, and
TestEveryLegWiresEveryRoleContract requires every leg to name or waive every
role contract. This PR adds three EpochFixture cases
(RunTokenSchemeChangeCarriesLiveAddressesAndOnlyThose,
RunACarriedAddressKeepsResolvingAcrossLaterBumps and
RunALostAddressStaysGoneAcrossASchemeChange); dolt, embeddeddolt and uow wire
them, and the http leg did not name them.

They are waived on the http leg with the existing
retentionEpochNoRoleWaiverReason, which already covers every other
RetentionFixture/EpochFixture case there: EpochFixture is raw storage hooks
(CurrentEpoch, BumpEpoch, MintUnderEpoch, StillServes, Resolve,
CurrentAddressFor, LoseVersion) with no publicops role behind them,
and internal/httpclient is a remote client with no epoch counter or address
table of its own. The constant's doc comment now counts eighteen cases.

Test-only; no production code changes. After this, the http leg's only
unnamed entrypoints are #6943's seven RunParticipationGeneration* contracts,
whose waivers belong to #6943.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(epoch): rename #6664's read-only store_epoch read past #6943's readStoreEpochInTx

#6943's "a create stamps participation even when the row is flagged
no-history" (7c00d65fa6) adds readStoreEpochInTx to version_history.go: the
mint's read of store_epoch, which SEEDS the singleton row when it is absent.
This PR's epoch_cas.go already declared a readStoreEpochInTx of its own, so
the package stopped compiling once #6664 was stacked on that commit.

The two cannot be folded into one. This PR's helper is deliberately
read-only: an absent row answers epoch 1 without writing, so
CurrentEpochInTx, StillServesInTx, ResolveEpochInTx, CurrentAddressForInTx and
LoseVersionInTx never write from a read (review item B4, pinned by
TestCurrentAddressForNeverReseedsAnAbsentStoreEpochRow and
TestCurrentAddressForWritesAndPublishesNothing). Calling #6943's seeding read
from those paths would break that.

So #6664's helper becomes readStoreEpochWithoutSeedingInTx, and its doc
comment says why it is not readStoreEpochInTx. The five call sites and the
ensureStoreEpochRow comment that names it follow. No behavior changes on
either side: #6943's mint path keeps its seeding read, and this PR's read-only
paths keep theirs. (ensureStoreEpochRow, which BumpEpochInTx and
MintUnderEpochInTx use, now has the same body as #6943's readStoreEpochInTx;
it is left as it is rather than folded in here.)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(storage): say exactly why the retention and epoch contracts stay waived on http

All but four of the participation-generation contracts (#6943's write
fence) now run over the served wire
(internal/httpclient/served_participation_generation_test.go): their subject
is a create or an update, which the client can send, and the fixture reads
the reference store's raw columns. The four stay waived, because the
mutation each needs has no form on the wire. This checked whether the
eighteen RetentionFixture/EpochFixture cases waived on the http leg,
including the three this PR adds, can be wired the same way. They cannot,
cheaply or otherwise, without a new wire operation:

- every EpochFixture hook (CurrentEpoch, BumpEpoch, MintUnderEpoch,
  StillServes, Resolve, CurrentAddressFor, LoseVersion) is the operation
  under test, not a seed or a raw read. Those operations are methods on the
  concrete *DoltStore and *EmbeddedDoltStore over issueops' *InTx bodies, on
  no storage interface, no publicops role and no wire operation. Every call
  path into them starts in a test: the methods are called only from each
  store leg's own _test.go files (its EpochFixture, which the conformance
  contracts drive, and its epoch tests), and the *InTx bodies only from those
  methods, from each other and from tests (the uow leg's fixture through a
  test-only type assertion, and issueops' and the dolt leg's own tests). No
  command, role or server handler reaches them, so no client mutation
  touches epoch state either;
- RetentionFixture's hooks (Resolve, Remove, Hold, ForceRemove, Erase, Mint)
  are implemented by no backend yet (every store leg wires them nil and its
  cases skip), and the wire has no removal, hold, erasure or gone/unknown
  resolve to bind them to.

Binding the epoch hooks to the reference store behind the served env would
pass, but as a property of that store, which the store legs already pin,
never of the client. So the waivers stay; retentionEpochNoRoleWaiverReason's
comment and reason string now say the above, in the shape #6661 gave
asOfReadNoHTTPWireWaiverReason, instead of the earlier general "raw storage
hooks with no publicops role" wording.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(epoch): report applied-but-unpublished epoch writes as indeterminate, bound address length, verify dup-key mints

Maintainer fixup for the #6664 review (adopt-pr attempt 1, iteration 1).

Major 1, applied writes reported as failed. An epoch writer whose SQL
commit has landed, or may have, now wraps storage.ErrCommitIndeterminate
on both legs and hands back what it applied, so nothing replays a bump
that has already advanced the counter:
- embedded BumpEpoch and BumpEpochCarrying return the new epoch with the
  error (failedBump), and MintUnderEpoch its address;
- server bumps wrap the sentinel on every Dolt failure after the SQL
  commit, including a failed staged-set read (unpublishedBump), and so
  does LoseVersion; an indeterminate SQL commit returns the epoch or
  address it may have committed, and a definite (typed MySQL) commit
  error returns neither.
Tests: TestEpochWritesReportAnUnpublishedCommitAsIndeterminate (new,
embedded, through an afterSQLCommit seam in runTransactionWithMessage);
on the server leg TestEpochWritesMarkAFailedStagedReadIndeterminate,
TestEpochWritesReturnWhatAnIndeterminateSQLCommitMayHaveApplied and
TestEpochWritesReturnNothingWhenTheSQLCommitFailsDefinitely, and the
sentinel asserted in every publication-failure test.

Major 2, address overflow. MintUnderEpochInTx refuses a pair whose
address would outgrow epoch_minted_addresses.address (VARCHAR(255)) at
the widest epoch the counter can reach, with a typed
ErrEpochAddressTooLong, so a carry can never fail on a lineage that
gained an epoch digit.

Major 3, duplicate-key mint. A duplicate key from the mint's INSERT is
no longer success: the row is read again and the mint stands only if
this transaction sees it served; a Gone row fails the mint with
ErrEpochAddressNotServed, an unreadable one with the duplicate-key
error. Measured on dolt-sql-server 2.2.0, the stale-snapshot INSERT does
not report a duplicate key at all: its COMMIT fails with Error 1213 and
the retried mint answers ErrEpochAddressNotServed
(TestAMintRacingTheMintAndLossOfItsAddressConflictsAtCommit).

Minors:
- absent store_epoch row: sqlmock cases pin that CurrentEpoch,
  StillServes, Resolve, CurrentAddressFor and LoseVersion never seed it;
  switching any of the five to the seeding reader now fails the test;
- the server publication tests mint their address instead of writing it
  by hand, and TestEpochCASTestsBuildEveryAddressWithEpochAddress scans
  that file too (issueops_test gains //internal/storage/dolt:go_srcs as
  data);
- stale text: the duplicate-key and carry comments, the issueops
  header's CompareAndSetVersionInTx citation, the embedded header's
  parity claim, and the 0072 header's "upserts the same row" (0072 is
  past the frozen floor).

Declined, for the owner:
- a scheme-change carry voids an older live lineage re-minted after a
  restore: rule (a), document and pin it, or rule (b), resolve the loser
  to the winner's carry, is a design call about R20-n, not a review fix;
- the uow epoch fixtures' in-transaction DOLT_COMMIT('-Am'): test-only
  fixtures with no production caller; routing them through RunTxScoped
  is follow-up work.

Validation: go build ./...; go vet; nogo (bazel --config=nogo) on
issueops, dolt and embeddeddolt; issueops epoch tests; the dolt epoch
tests against a real dolt-sql-server (56 pass) and the fake-backed
publication tests; embedded epoch tests with BEADS_TEST_EMBEDDED_DOLT=1;
uow epoch contract; the schema package; the address guard under bazel
in the local sandbox. All 14 overlay mutants of the fixes are killed.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Jim Wordelman <mayor@quad341.uk>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Julian Knutsen <julianknutsen@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-18 21:20:57 +00:00
Jim Wordelman
9c6f6e15cd
bd versions + a config switch for versioned history — the minimum cut that makes versioned beads usable (#6138) (#6661)
* test(conformance): author R7.1 as-of-read conformance contract

Adds a new Phase 0 conformance contract for R7.1 (as-of reads,
gastownhall/beads#6136, design doc #5898 rev 9) -- no fixture existed for
this anywhere in backend/conformance/, so this is authored from scratch,
not un-skipped.

AsOfReadFixture supplies three hooks: Resolve (the read under test),
MintAt (seeds a version at a controlled accept-time), and
MakeUnanswerable (arranges a Gone/Unknown refusal independently of
RetentionFixture/EpochFixture's own removal machinery, since R7.1's
refusal is a sibling contract to R17/R20, not a dependent of them).

Five Run* functions cover the bead's five required-coverage points
verbatim:
- selector accepts either an instant or a version address, and both name
  the same answer when they resolve to the same version;
- resolution returns the latest version at-or-before T, with a version
  accepted after T strictly excluded even when close by;
- a served result reports the exact address it served;
- an unanswerable read refuses as a typed Gone*/Unknown outcome with a
  reason, and never substitutes a neighboring version, the nearest
  surviving version, or current state -- asserted directly by requiring
  Address == "" on every refusal;
- local completeness only (TC3): a lineage minted into one of two stores
  resolves there and refuses Unknown (FR-06) on the store that never
  received it, asserting nothing about agreement between the two.

The file's doc comment states the Memory-plane-consumer-policy non-goal
explicitly (closed-before-T boundary, sibling exclusion, supersession
exclusion are consumer policy per #5877, not part of this primitive) so
the gap reads as deliberate scope, not an oversight.

Wired onto all three storage legs (dolt, embeddeddolt, uow) with
unconditional "not implemented" stubs rather than partial-real hooks, so
every case fails via MintAt on its first fixture call -- deterministic
RED with no risk of an accidental pass, per the precedent set by
be-x5jqd.1. Sentinel error messages reference be-x5jqd.5, the
implementation child this suite is written ahead of.

TestEveryLegWiresEveryRoleContract requires zero waivers: all 5 new
entrypoints are fully wired (not nil) on all three legs.

Branch cut from the rebased be-2h167 head (5fdfb92fe5).

Verified via go build ./..., go vet ./..., gofmt -l, and go test -tags
cgo across all three legs: all 5 new subtests (8 counting the 4-way
Gone/Unknown table) FAIL for the documented reason,
TestEveryLegWiresEveryRoleContract passes cleanly on all three legs, and
TestVersionedHistoryPhase0RunsInFullSkipMode is unaffected.

Bead: be-x5jqd.2

* fix(history): refuse metadata integers binary64 cannot hold exactly (refs be-wdlod, #6358 follow-up)

canonicalDurableState canonicalizes numbers per RFC 8785 (JCS), which
renders every number as an IEEE-754 binary64 double regardless of
magnitude. An integer-valued metadata literal past 2^53-1 therefore
rounds to a DIFFERENT VALUE on the way into durable_state -- not a
formatting change, a value change -- which can collapse two distinct
issue states onto one durable_state and therefore one content token.

refuseUnrepresentableIntegers is the fix: an admission gate, one layer
above the JCS step, that parses each JSON number literal exactly with
math/big (never float64 -- float64 rounding is the exact thing being
guarded against) and refuses whenever the literal's VALUE, not its
spelling, is an integer whose magnitude exceeds 2^53-1. Value-not-form
matters because the first cut of this gate skipped any literal
containing '.', 'e', or 'E' outright, reasoning that non-integer forms
round-trip losslessly -- true for genuinely fractional literals, but
not for an integer VALUE spelled in decimal or exponential form: QA
found that 9007199254740993.0 and 9.007199254740993e15 walked straight
past that first gate and canonicalized to the same colliding target as
the bare literal. The predicate now inspects the value via
big.Rat.SetString/.IsInt, so all three spellings are refused alike.

1e300 is refused too, no notation carve-out: it denotes the integer
10^300, and binary64 spacing near that magnitude is roughly 2^248 --
vastly wider than the collision window at the 2^53 boundary this gate
exists to close. Admitting that spelling while refusing 9007199254740993
would be the same defect with a bigger window (mayor's ruling, citing
this collision-window argument as decisive). Non-integer values stay
admissible past the bound (1.5e300 is fine): ES6 shortest-round-trip
formatting is a bijection on doubles, so two non-integer literals
collide only if they already denoted the same double before this gate
ever ran, and a fractional literal never carried a promise of exact
decimal storage the way an integer -- an id, a count, a timestamp --
carries a promise of exactness that silent rounding would corrupt.

jcsCanonicalize is canonicalDurableState's RFC 8785 step pulled out on
its own, below the gate rather than behind it, because refusing 1e300
puts the admission policy at odds with one of this store's own spec-
conformance fixtures: testdata/jcs/input/values.json, one of the 10
official RFC 8785 vectors shipped verbatim from the spec's reference
suite (not authored here), itself contains 1E30. RFC 8785 conformance
is a claim about the canonicalizer, not about this store's admission
policy, so TestCanonicalDurableStateIsJCS now runs all 10 shipped
vectors -- values.json included, still asserting 1E30 folds to 1e+30
per spec -- through the ungated jcsCanonicalize, while its two local ad
hoc cases keep exercising the full gated canonicalDurableState, since
neither conflicts and both are more faithful to production traffic that
way. The gate's own refusal of that magnitude is pinned separately, on
its own terms, in version_history_integer_admission_test.go.

Severity is MINOR, not MAJOR: Dolt's JSON column type already
renormalizes metadata to float64 on every read regardless of this gate
(measured on dolt 2.2.3: 9007199254740993 reads back as
9007199254740992, 1e300 as 1e+300), which converges two distinct
in-memory states to one stored state before any version row is minted.
The collision this gate closes is therefore unreachable through normal
bd update traffic on this backend; it stays real for off-path writes --
bd sql, a hand-written JSONL import, or a TEXT-column backend -- which
is why this landed as a follow-up rather than gating Phase 2 itself.

Verified clean: a full module compile; static analysis over the
storage tree; the complete issueops package suite (260 subtests); the
schema package suite; and the broader storage tree, apart from one
dolt-package test that fails identically against the pre-this-diff
baseline (a Dolt server port collision with a concurrent session's
process on this host) -- confirmed pre-existing and environmental, not
caused by this change.

* feat(issueops): implement R7.1 as-of read's store-layer primitive (be-x5jqd.5)

Turns backend/conformance/versioned_read_contract.go (authored RED in
6aefdce60, gastownhall/beads#5898 rev 9, #6136) towards GREEN. This
commit lands the shared production seam; wiring the three per-leg
fixtures (dolt, embeddeddolt, uow) onto it is a follow-up commit.

asof_read.go (new): AsOfReadInTx resolves an AsOfSelector (an instant
or a previously-served version address) to one issue_versions row --
the latest version accepted at or before T, everything later strictly
excluded -- and reports either its state and address, or a typed
refusal (gone-retention/gone-erasure/gone-reorganization/unknown) with
a reason. storeID is accepted only for signature parity with the
fixture hook shape (Design D: the conformance suite's own ad hoc
dispatch string selects among per-leg isolated test stores; issue_versions
carries no store_id column, matching CurrentEpochInTx's precedent).
AsOfRestriction is string-backed, unlike R20's int-backed
EpochRestriction, because it round-trips directly through the new
removed_restriction column with a plain cast.

Migration 0069 adds issue_versions.removed_restriction VARCHAR(30),
nullable (NULL = live, never a stored value). Sits alongside removed_at/
removed_reason (migration 0067, virgin until this slice); AsOfReadInTx
only READS all three -- a future R17 is the real writer via hold/remove/
erase primitives, no collision. Guarded PREPARE + CLI direct-DDL
override + schema_test.go hygiene entries, all modeled on 0068's
pattern; TestLatestVersionIncludesMigration0068 is superseded per the
one-pin-at-a-time convention.

version_history.go: splits RecordVersionInTx into a versionedHistoryEnabled-
gated wrapper and a new exported RecordVersionAtInTx, both delegating to
a shared recordVersionAtInTx(..., at time.Time) body. RecordVersionAtInTx
is test-support only -- ungated, caller-chosen timestamp -- so an as-of-
read fixture's MintAt hook can mint version rows dated in the past
without flipping a store's history-enabled flag. No behavior change for
existing RecordVersionInTx callers.

The new/moved code shifted which function each completeness guard sees
as the enclosing write site, so all three needed updating, not bypassing:
- row_lock_guard_test.go's funcNameExemptions key moves from
  RecordVersionInTx to recordVersionAtInTx (the write's new actual
  enclosing function), reasoning extended to cover both callers.
- journal_completeness_test.go's beadDMLExemptions and
  version_completeness_test.go's versionExemptions each gain a
  RecordVersionAtInTx entry mirroring RecordVersionInTx's own (test-
  support twin, no production caller, no mutation of its own to
  journal/mint beyond the seam itself).

Also fixed a latent test-helper gotcha found while writing the 0069
Dolt-CLI test: queryDoltCSV trims the whole `dolt sql -r csv` output
before parsing, so a row whose ONLY selected column is NULL renders as
a bare blank line that strings.TrimSpace silently swallows when it's
the trailing line, collapsing "one row, NULL value" into "zero rows".
Fixed by selecting a non-NULL anchor column (issue_id) alongside the
column under test.

Verification performed: full module compile and static analysis, gofmt
clean; the complete internal/storage/schema suite with the real-Dolt-CLI
lane enabled; and the complete internal/storage/issueops suite,
including every completeness/row-lock structural guard -- all clean.

* feat(issueops): wire dolt, embeddeddolt, and uow onto R7.1 as-of read (be-x5jqd.5)

Turns backend/conformance/versioned_read_contract.go fully GREEN
(authored RED in 6aefdce60, gastownhall/beads#5898 rev 9, #6136): this
follows up on 4da98a9a8's shared production seam by wiring all three
storage legs' AsOfReadFixture (Resolve/MintAt/MakeUnanswerable) onto
issueops.AsOfReadInTx, per that commit's own stated plan.

Each leg's asOfReadHarness is Design D's dispatch table: the
conformance suite invents storeID strings to select among N isolated
stores, and isolating them is entirely the harness's job --
AsOfReadInTx itself stays storeID-oblivious. The three legs differ
only in how they get that isolation cheaply:
  - dolt: one shared testcontainer, N randomly-named databases
  - embeddeddolt: N fully independent on-disk databases
  - uow: one shared dolt sql-server subprocess (NewDoltServerUOWProvider's
    adoption semantics collapse repeat calls against the same root dir
    onto the same process), N randomly-named databases

All three harnesses are Fatalf/Skip-free below their own
TestAsOfReadContract body: AsOfReadFixture's hooks carry no *testing.T
of their own and run from inside t.Run subtest closures (sometimes
nested two deep), where FailNow/SkipNow would panic with "subtest may
have called FailNow on a parent test". Failures instead flow back as a
plain error into conformance.RunAsOfReadXXX, which Fatalfs correctly
on the subtest's own goroutine.

Two small companion fixes surfaced while wiring the legs on:

  - issueops.asOfVersionAddress -> AsOfVersionAddress (exported): each
    leg's MintAt needs to build the exact token its own
    RecordVersionAtInTx write will resolve against, from a different
    package than issueops.

  - recordVersionAtInTx now truncates `at` to whole seconds before
    writing it to issue_versions.change_at (DATETIME, migration 0067).
    Measured on dolt 2.2.3: the column ROUNDS a sub-second value rather
    than truncating it, so a change_at of 13:32:04.77 lands at 13:32:05
    -- strictly later than the instant it was meant to record. That
    breaks resolveAsOfRevisionInTx's "at or before T" comparison
    (asof_read.go) for any T between the true instant and the rounded-
    up one. This is a production correctness fix, not test-only:
    RecordVersionInTx hits the identical hazard on every real
    time.Now().UTC() call whose nanosecond component rounds up.

Verification performed:
  - uow: TestAsOfReadContract green standalone (5 subtests, 4 nested
    restriction variants) and the full uow package green in isolation
    (673s, includes pre-existing TestReleaserContract and everything
    else). A `go test ./internal/storage/...` full-tree sweep flagged
    uow FAIL via a 600s hang inside TestReleaserContract's migration
    lock -- traced to cross-package parallel resource contention from
    running the whole storage tree's dolt-sql-server-spinning suites
    concurrently, not a regression: the same test is clean when the
    package runs alone.
  - embeddeddolt: TestAsOfReadContract green (22.89s, all subtests and
    all 4 restriction variants).
  - dolt: skipped on this host -- no Docker for the leg's testcontainer,
    the same pre-existing skipIfNoDolt gate the whole dolt package
    already runs under, unrelated to this change.
  - TestEveryLegWiresEveryRoleContract and
    TestEveryContractImporterInTheStorageTreeIsRegistered both pass for
    all three legs.
  - gofmt clean; whole-module build and vet clean.

Bead: be-x5jqd.5

* feat(versions): bd versions + a config switch for versioned history (refs be-hs42e.6.2)

Versioned history has been buildable but not usable: migrations 0067+ create
issue_versions, and the Phase 2 writer fills it, but SetVersionedHistoryEnabled
was a Go method nothing in cmd/bd ever called. There was no supported way for a
person to turn the feature on, and no way at all to read it back without raw
Dolt SQL. This adds the switch and a reader.

- versioned-history.enabled (env BD_VERSIONED_HISTORY_ENABLED), default off,
  applied at the top of wireStorageDecorators. It MUST run on the raw store:
  the capability is a type assertion and no decorator forwards it, so applying
  it after a wrap would silently leave history off while the config said it was
  on. TestVersionedHistoryConfigAppliesToTheRawStore pins the ordering, and
  fails rather than skips if a decorator ever starts forwarding.

- storage.VersionLister + issueops.ListVersionsInTx over issue_versions, on the
  dolt and embeddeddolt legs. A narrow capability rather than a Storage method,
  following ExternalRefHistoryQuerier: a backend that cannot know must be able
  to say so instead of answering "none".

- bd versions <id>, with three distinguishable outcomes rather than two. Off,
  unsupported, and empty are different answers with different fixes; an empty
  list is a claim about the bead, and returning it for the other two would be a
  wrong answer rather than an empty one. The four removal restrictions render
  as four sentences, and unknown (no lineage knowledge here) is not gone.

- REV is printed as a local ordinal with a footer saying so. It is deliberately
  NOT presented as a citable address: AsOfVersionAddress mints over the ordinal,
  two clones can both hold revision 8 of the same bead, and the durable
  version_id primary key has not landed. An address printed now would break
  every citation once it does.

- bd version <id> now points at bd versions <id> instead of printing bd's own
  build and looking like the bead has no history. Bare bd version is unchanged.

No migration: 0067/0068/0069 already carry every column this reads.

Verified on a scratch store: create + 2 updates + close with the flag on lists
4 attributed versions; with the flag off, 0 rows in issue_versions and the
command refuses naming the key.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(versions): emit local_revision, not revision, in --json (bee's #5898 consumer read)

types.Issue already ships `json:"revision"` (types.go:1206) and it is the
row-lock CAS token — an unrelated value, and a string where this one is an
int64. Emitting "revision" from the versions payload would put two meanings
and two types on one key in the place machines read: the `bd version`
collision again, with worse consequences, since a script that learned one
shape gets the other and cannot tell.

local_revision also carries the non-citability with the value rather than
only in a footer no script parses, and needs no rename when version_id lands
beside it.

Raised by @bee-ghosttrack's consumer read on #5898, who reported a couple of
hundred script files parsing `bd … --json` — this is the part of the surface
they would actually touch. Pinned by
TestVersionsJSONDoesNotCollideWithIssueRevision.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(versions): make the advertised switch actually switch it on (#6661 review)

@bee-ghosttrack's blocking finding: the help text and the off-refusal both
told the user to run `bd config set versioned-history.enabled true`, and that
command did not turn the feature on. The key is not yaml-only, so `bd config
set` routed it to the workspace DATABASE; both readers called config.GetBool,
which is viper (yaml + env) only. The user ran exactly what they were told,
was told it was set, and recording stayed off — configured on, actually off,
one layer up from the decorator-ordering hazard this command already guards.

Takes bee's option (b): read the switch from the store's own config, where
`bd config set` already writes it. That also makes recording a property of
the STORE rather than of whoever is calling, which is what closes finding 2 —
two clients of one sql-server can no longer disagree about whether a write is
recorded.

The two planes are OR'd rather than ranked. The hazard is a write that fails
to record, never one that records when it needn't, so either saying yes is
enough and neither can silently switch the other off. BD_VERSIONED_HISTORY_ENABLED
still works for a one-off run. A store with no settings plane (proxied, no-db)
falls back to viper instead of failing the command.

Also from that review:

- finding 3: the read no longer refuses just because recording is off. A store
  that recorded for a month and was switched off HAS versions; refusing made a
  claim about the store from a fact about this invocation's config, which
  VersionLister's own doc says are independent. Rows are listed either way,
  with a note when recording is currently off; refuse-with-remedy only when
  off AND empty.
- finding 4: an unresolved id with no versions is "no such bead", not "none
  yet". Falling through on an unresolved id stays right — a deleted bead can
  still have versions — but only until the answer is empty.
- finding 5: the ordering test's escape hatch is t.Fatal, not t.Skip; a skip
  returned before the assertion that actually guards the ordering. The
  storage_chain.go doc cited a test name that never existed. version_lister.go
  promised a list of compile-time assertions that was not there; the
  assertions now exist in each leg's own package, where they can compile.
- `versioned-history.` is a recognized config prefix, so the set no longer
  prints "not a recognized config key" immediately before confirming the set.

Test: TestVersionedHistorySwitchRoundTrip_Embedded drives the real
`bd config set` through to a real `bd versions` against a real store, with
BD_VERSIONED_HISTORY_ENABLED cleared so a pass proves the store setting
carried it alone. This is the round trip bee asked for, and it is the only
shape that could have caught the original defect — a fake-backed unit test
cannot see a disagreement between two real planes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* docs(versions): name what the OR'd switch planes cost (#6661 review)

@bee-ghosttrack, round 2: with the two planes OR'd rather than ranked, the
consequence worth naming is that once the store setting says on,
BD_VERSIONED_HISTORY_ENABLED=0 cannot countermand it for a single run. The
help now says so, and says where to turn it off instead.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(versions): gate the wiring store read on the capability, not a recover (#6661 review)

@bee-ghosttrack's round-3 blocking finding. wireStorageDecorators reads the
versioned-history switch off the raw store on every bd invocation, and did so
before anything established that the store could use the answer. cmd/bd's own
fakeProbingDoltStore embeds a nil storage.DoltStorage, so the promoted
WorkspaceConfig SIGSEGV'd at wiring time -- and a panic takes the whole test
binary down, which is why one unguarded read showed up as two red CI steps
(Test (macos-latest), PR Core (wrapper timing)) rather than one failed test.

Takes the fix bee prescribed: assert storage.VersionedHistoryConfigurer first
and read the store plane only when the store could act on the answer.

The previous round fixed this with a recover plus a WorkspaceConfig on the
fake, and left a comment arguing that no capability assertion could work. That
argument was wrong, and wrong in a specific way worth recording: it reasoned
about asserting WorkspaceConfig, which IS declared on storage.DoltStorage and
so matches every store including the nil-embedding fake. bee proposed
asserting VersionedHistoryConfigurer, which is a different interface --
SetVersionedHistoryEnabled is NOT part of DoltStorage, so the fake does not
promote it and does not satisfy it. The comment rebutted a fix that had not
been proposed. It now states what is actually true.

So the fake goes back to byte-identical with the pushed head: the wiring fix
fixes that test without touching it, as bee said it would, and the fake keeps
the "only the methods below may be called" constraint its own doc asserts.

The recover stays, as a backstop rather than the guard. It covers what the
assertion does not: versionedHistoryEnabled is also called from `bd versions`,
where the read is the point and no capability gate precedes it, and a store
that DOES implement the capability can still have a half-built settings plane.

The viper plane is still read first and unconditionally, so a set-but-unsupported
store still reaches applyVersionedHistoryConfig and still gets its warning. Only
the store-plane read is gated. The one case that loses a warning is a store whose
own settings plane says true but which lacks the capability -- and `bd config
set` could not have written that plane on such a store to begin with.

Tests, split so each layer fails for its own reason:

- TestWireStorageDecoratorsSkipsTheStoreReadForANonConfigurer asserts calls == 0,
  not merely that it survived. "It survived" is what the recover delivers too,
  which is how this regressed the first time.
- TestWireStorageDecoratorsSurvivesAStoreThatCannotAnswer keeps the recover's
  case, now on a store that IS a configurer so the guard lets it through.
- TestVersionedHistoryEnabledForWiringReadsTheStoreWhenItCanRecord pins that the
  guard did not stop reading the plane finding 1 existed to start reading.

Mutation-verified both ways: removing the guard fails the first test on
calls=1/want=0 while everything still "passes" by surviving; removing the
recover leaves the first test green and fails only the second.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* test(versions): name the embedded round-trip so CI selects it (#6661 review)

@bee-ghosttrack's round-3 finding 2. The round-trip test that proves the two
config planes agree ran only on the author's machine.

.github/scripts/embedded-test-shard.sh discovers what to run with
`grep -rh '^func TestEmbedded' cmd/bd/*_embedded_test.go` and validates
manifest entries against ^TestEmbedded[A-Za-z0-9_]+$.
TestVersionedHistorySwitchRoundTrip_Embedded matched neither, so the only
lane that sets BEADS_TEST_EMBEDDED_DOLT=1 for cmd/bd never selected it; the
generic Test jobs run without that env, where the test's own guard skipped it.
Skipped and passed look alike from outside, which is what made it invisible.

Renaming is the whole fix -- a test absent from the shard manifest still gets
a shard from the fallback distribution. Verified by running the script's own
selector over the tree and confirming the new name is discovered.

The rationale is now a comment on the test, because the old name reads more
naturally and nothing else in the file explains why the prefix is load-bearing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* test(versions): drop the dead fixture, name the empty arm (#6661 review)

@bee-ghosttrack's round-3 finding 3. TestVersionsRefusesWhenFlagOff built a
backend holding one recorded version, handed it to nothing, and ended in
`_ = backend`.

The fixture was load-bearing for the assertion the test was named after --
that the read was refused BEFORE the backend was consulted. Finding 3 removed
that behaviour on purpose, so the assertion went with it; the fixture stayed
behind, unused, under a name still promising the old contract.

Renamed to TestVersionsOffAndEmptyRefusesWithRemedy for what it actually pins,
moved beside TestVersionsOffButRecordedStillLists so the two read as the pair
they are -- recording off decides the FOOTER, never whether to read; off with
rows lists them, off and empty is the one case that refuses -- and the dead
fixture is gone.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(proxy): give bd versions a capability row, a typed refusal (#6661)

main's proxied-server capability registry (#6582) requires every command
path to carry a row, and TestProxyCapabilityRegistryCoversCommandTree
fails for `bd versions`, which this PR adds.

It is refused, reason unimplemented, rather than permitted. It is a pure
read (issueops.ListVersionsInTx in one read transaction), so nothing
about it is unsafe on a shared backend and "design" would be wrong. But
it reaches the backend only through the optional storage.VersionLister,
which the dolt and embedded stores implement and the uow provider does
not. Permitting it would open the proxied provider in the root pre-run
(starting the proxy and dolt) and then fail untyped at
errVersionsUnsupported. The row refuses with proxy.versions.unsupported
before any side effect, and its tracking names the missing piece: a
VersionLister over the proxied uow provider (tracker #6132).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(uow): register verified shutdown cleanup in TestAsOfReadContract (#6661)

main's TestEveryServerFixtureRegistersVerifiedCleanup (#6441) requires
every server fixture to use verifiedShutdownCleanup(t, storeRootDir)
instead of calling proxy.Shutdown directly, so a shutdown that refuses
cannot leave a live dolt sql-server behind a green run. TestAsOfReadContract
still had its own proxy.Shutdown cleanup that only logged the error.

The cleanup is registered at the same point, so the LIFO ordering the
harness relies on (every provider closed before the shared server stops)
is unchanged.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* style(schema): gofmt cli_migrations.go after the 0070 renumber

The rebase onto upstream's 0069_widen_issue_versions_datetime_precision
(#6675) joined its const and the renumbered 0070 comment without the
separating blank line, which make fmt-check rejects.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(history): stop flooring change_at now that 0069 widened it (refs be-hs42e.10)

recordVersionAtInTx carried `at = at.Truncate(time.Second)`. That was correct
while 0067's plain DATETIME was in force: Dolt's datetime(0) does not truncate
sub-second input, it ROUNDS half-up, so a change_at could be stored strictly
LATER than the instant it recorded and break AsOfReadInTx's "latest version
accepted at or before T". Flooring first left nothing to round.

Migration 0069 widened change_at and removed_at to DATETIME(6), which removes
the rounding entirely and makes the floor actively harmful — it threw away the
microseconds the widening exists to keep. Two versions minted in the same
second collapsed onto one stored value and stopped being distinguishable by
change_at, which defeats --at, the only PORTABLE selector the read surface
offers while revision stays a per-store ordinal and no durable version address
has shipped.

This was measured, not inferred. On a store carrying both the widening and the
floor, four versions minted inside one second all read back 15:40:38.000000:
the column could hold microseconds and the writer refused to supply them. Two
correct fixes cancelling each other, and only the second one was visible.

removed_at needs no matching change: MarkLatestVersionRemovedInTx already
passes time.Now().UTC() straight through and was never floored. Audited rather
than assumed.

Pinned by TestRecordVersionKeepsSubSecondChangeAt (embeddeddolt), which mints
two versions 500ms apart inside one second and asserts both that the
fractional part survives and that the two differ. Verified it FAILS with the
floor restored — reporting exactly the 11:04:07.000000 collapse — and passes
without it, so it is a real guard rather than a test that passes either way.
The column's own half of the contract stays pinned by
TestMigration0069ChangeAtSurvivesSubSecondPrecisionThroughDoltCLI.

Also drops this branch's own tx-minting guard commit, which upstream landed
independently as #6833.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* test(feat): red — Activate versioned history in every store and provider factory, each store reading its own #6661 switch (stacked on #6661) (refs be-qxq70)

Tests first, all failing on the right reasons at fb6e219d6a (#6661's head):
- D1: on real embedded stores a routed update, a routed close and
  `bd create --repo` into a rig whose OWN row is on each gain 0
  issue_versions rows, want 1; the control (a write inside the rig) gains 1.
- D3: BD_VERSIONED_HISTORY_ENABLED=1 records into no routed store, and =0 is
  shown not to switch off a store whose row is on.
- D4: TestEveryStoreConstructionActivatesVersionedHistory flags all 9 store and
  provider factories, and its mutation probe reports "probed 0 of 9".
- wiring: wireStorageDecorators still reads the store's row (1 read) and
  switches it on; activation belongs to the factories.
- internal/versionedhistory (does not exist yet) and the doctor repair opener:
  the resolver, the provider read (K1), the error-vs-absent diagnostics and the
  shared-activation requirement.

* feat: green — Activate versioned history in every store and provider factory, each store reading its own #6661 switch (stacked on #6661) (refs be-qxq70)

Stacked on #6661 (fb6e219d6a). #6661's plane and approved commits are
unchanged: the switch stays a row of the store's own config table, OR'd with
viper, and `bd config set versioned-history.enabled true` still writes that
row and not config.yaml.

The defect: applyVersionedHistoryConfig ran only from wireStorageDecorators, so
the one store a command opens for its own workspace was activated and every
store built by newDoltStoreFromConfig (prefix-routed update/close, `bd create
--repo`, the direct-mode open) was not. With the switch on they recorded no
issue_versions row while the command reported success.

- Activation moves to where a store or unit-of-work provider is constructed
  (the store_factory.go/_nocgo.go twins, uow_factory.go's two provider
  constructors, the personal-migration planning store, and bd doctor's repair
  opener), on the raw store, before anything can wrap it. wireStorageDecorators
  no longer activates, so each store's row is read once per open.
- One resolver, in internal/versionedhistory, reachable from doctor/fix which
  cannot import package main. #6661's semantics are preserved: one key
  spelling, viper first, then the capability check, then the row read, with
  the recover backstop. The behavior-preserving refactor of #6661's
  versioned_history_switch.go / storage_chain.go is the only edit to files
  #6661 introduced; the resolver tests moved with the code, assertions
  unchanged.
- K1: a provider has no DoltStorage, so its row is read through a unit of work
  (NewUOW + ConfigUseCase().GetConfig) at construction: a pooled connection,
  one SELECT, rolled back. The constructors have already run initSchema, so
  this starts no server and creates nothing. It does not fall back to viper
  alone.
- A row read that errors resolves to off but is reported under BD_DEBUG; an
  absent row stays silent.
- Docs (configuration.md), a CHANGELOG entry, and a hint at `bd config set ...
  true` time that the setting replicates and is single-writer only.

The bazel-sync run also wanted unrelated BUILD entries for files that already
exist at #6661's head (versions.go, versioned_read_contract.go, ...); they are
not part of this change and are left out.

* test(feat): green — a served mutation records from the provider's own row, on a real server (refs be-qxq70)

K1 is the one place the resolver cannot ask a DoltStorage: `bd serve` against a
server-mode workspace writes through a unit-of-work provider, so its flag is
read through NewUOW + ConfigUseCase().GetConfig at construction. The fakes in
internal/versionedhistory prove the read; this proves it WORKS against a real
Dolt server at that moment (pool open, database selected, schema migrated).

Two server-mode workspaces on the shared test container: the one whose own row
says on records a served HTTP mutation (+1 issue_versions row), the one with no
row records none.

Measured both ways: passes with the provider activation in place, and fails
("added 0 issue_versions row(s), want 1") with it deleted from both provider
constructors in uow_factory.go.

Proxied lane only (BEADS_TEST_PROXIED_SERVER=1 BEADS_TEST_ENV_RUN_DOLT=1
TESTCONTAINERS_RYUK_DISABLED=true), like TestServeActivatesTheEventsJournal, so
it skips in a plain make test.

* test(feat): red — #6661 review round, small items: findings 2, 4, 5, 6 and 8

Tests for the small findings in the second review round on #6661.

- 2: `bd versions` must decide "recording" from the store's own setting.
  BD_VERSIONED_HISTORY_ENABLED=1 with the store's row unset and nothing
  recorded must refuse, not answer "none yet"; the help text must not offer
  the variable as a per-run switch for a read.
- 4: `bd versions --json` is one object, {"versions": [...], "recording":
  bool}, with [] rather than null when nothing is recorded.
- 5: AsOfReadInTx reports any stored removed_restriction outside its four
  values, "live" included, as unknown.
- 6: the 0070 header spells the restriction vocabulary with the underscore
  spellings the code stores and compares.
- 8: `bd version <arg>` exits 1 and points at `bd versions <arg>`. This one
  pins behaviour the branch already has, so it passes at red; it was shown
  to fail with the guard removed.

Item 7 is a comment-only fix and has no test.

* feat: green — #6661 review round, small items: findings 2, 4, 5, 6 and 7

- 2: `bd versions` reports the store's own setting. It read the environment
  and config.yaml as well, so BD_VERSIONED_HISTORY_ENABLED=1 made a store that
  never recorded answer "No versions recorded yet" instead of refusing with
  the remedy. It only reads, so those planes cannot make the store record.
  The "for one run" hint is dropped from the help and the refusal, since it
  could only ever produce the wrong answer.
- 4: `bd versions --json` is always {"versions": [...], "recording": bool},
  with [] rather than null when nothing is recorded. Previously a bare array,
  or null, and no recording flag.
- 5: AsOfReadInTx reports a stored removed_restriction outside its four
  values, "live" included, as unknown instead of passing it through.
- 6: the 0070 header and its test use the underscore spellings the code
  stores and compares. 0070 is unreleased, and this changes its content hash.
- 7: comment only. DATETIME(6) still rounds to the microsecond; the comment
  said 0069 removed the rounding entirely.

* test(fix): red — #6661 number admission: fractions past the bound, exponent cost, pre-enable scan

Tests for the review round on the number-admission gate, written before the change:

- a fraction whose nearest double is at or past 2^53 is refused (9007199254740993.5,
  9007199254740994.5, 9007199254740991.5), and the gate is closed under its own
  canonicalization, checked as a property over a corpus
- classification depends on magnitude, not on the size of an exponent: 1e100000000 and
  -1e400 are refusals, 1e-100000000 and 1e-400 are admitted, and allocation does not grow
  with the exponent
- the sentinel's text names the I-JSON exact-integer range
- CheckMetadataVersionable agrees with the mint, and FindUnversionableMetadata finds what
  the mint would refuse and skips what it never versions
- bd config set versioned-history.enabled true refuses, writing nothing, over rows the mint
  would refuse, and false never scans

Guards that pass before and after by design: the verdict on every literal as the store
returns it, the mint's refusal writing nothing, and, on the server and embedded legs, the
store-level property that no two literals reading back as different numbers share a canonical
form. Those store tests also log what each Dolt holds for a number outside the range when
history is off, and pin that a later write to such a row is refused once history is on.

* fix: green — #6661 number admission: the nearest-double rule, ParseFloat classification, a pre-enable scan at the switch

The admission gate now refuses a JSON number literal if and only if the binary64 nearest to it
has magnitude above 2^53-1, or it overflows, whatever the literal's spelling:

- a fraction that rounds to or past 2^53 (9007199254740993.5, 9007199254740994.5,
  9007199254740991.5) is refused, so the gate is closed under its own canonicalization
- literals are classified with strconv.ParseFloat, linear in their length, so 1e1000000 no
  longer builds a million-digit integer inside the writer's transaction; an overflow is the
  sentinel and a magnitude too small for binary64 is admitted; math/big is gone
- the sentinel reads "number is outside the I-JSON exact-integer range (magnitude above
  2^53-1)"; ErrIntegerNotRepresentable keeps its name, with a comment saying why
- ordinary fractions such as 0.1 stay admitted; no verdict on a value the store hands back
  changes

The header now says what the mint's refusal is: an integrity backstop inside the writer's
transaction, not an admission policy for metadata.

bd config set versioned-history.enabled true now checks the store first. It reads every issue
the mint would version, runs the mint's own function over each one's metadata
(issueops.CheckMetadataVersionable, through issueops.FindUnversionableMetadata), and refuses,
writing nothing, while any is refused, printing the count, the ids and the fix. There is no
override; setting it false never scans. The environment and config.yaml planes do not pass
through the command and rely on the refusal at write time.

Also: docs and a CHANGELOG entry, and comments in the existing admission tests that described
the old rule.

* test(fix): keep the exponent-cost allocation test stable under the race detector

TestNumberAdmissionAllocationsDoNotGrowWithTheExponent compared two
independently averaged allocation counts with a strict >. Under -race
sync.Pool drops one Put in four at random, and the refusal's fmt.Errorf
takes its printer from a pool, so a call now and then allocates one more
than it would. The test failed on most runs of `go test -race -short`, the
flag set of the required lane, and never without -race.

That noise only ever adds allocations, so compare the fewest allocations
over many single-call samples instead. The comparison stays strict, and it
still fails on the exact big.Rat gate this test exists to keep out (49
allocations for 1e1000000 against 36 for 1e400). The doc comment no longer
claims the count is deterministic under -race.

* test(fix): red — set-many skips the pre-enable scan; pin what the scan reads

bd config set-many versioned-history.enabled=true switches versioned history
on with no pre-enable scan: only bd config set makes it. With a row holding a
number outside the I-JSON exact-integer range the setting is turned on and the
next write to that row fails. The new embedded test drives the real verb over
such a row and expects the whole batch to be refused, writing nothing.

The scan's own embedded test only used open issues, so deleting AllFlag or
IncludeAllTypes, or replacing the unlimited Limit, left it green. It now adds a
closed, a pinned and a gate offender, and a second test imports more rows than
one default page and expects every one of them counted.

* fix: green — bd config set-many runs the pre-enable scan for versioned-history.enabled

bd config set refuses to turn versioned history on while an issue holds
metadata a version could not record. bd config set-many writes the same
setting through a validation loop of its own and made no such check, so it
switched history on over a row whose next write then failed. It now runs the
same check for that one key inside its validation loop, before any pair of
the batch is written, so a refusal stores nothing. The check reads through
the accessor it already used on both routes: the store the command has
opened on the direct route, the provider on the proxied one.

* chore: strip internal tracker ids from the comments and tests across #6661's range

Comment and test-text changes only; no behaviour changes. Across every commit from the
merge-base with upstream main to here, the comments and tests named internal tracker ids
(one bead id in eleven comments and in one test failure message, an example id in the
`bd version` guard's comment, and an internal decision's attribution in the huge-exponent
admission test). They now cite gastownhall/beads#5898 and #6136 only, which is what a reader
of the public branch can follow.

The 0070 migration's header comment is one of the edits. That migration is new on this branch
(it does not exist on upstream main), so the change is not an edit to a shipped migration, but
it does change the file's content hash; the schema tests that compute hashes are self-consistent.

Text that already exists on upstream main is left as it is, even where it mentions an id: this
commit does not rewrite lines the range did not add.

* chore(bazel): bazel-sync (gazelle) for the #6661 range

* fix(bazel): give fix_test the store.go its activation guard parses

TestVersionedHistoryDoctorRepairOpenerUsesTheSharedActivation parses
store.go from its working directory. `go test` runs in the package
directory, so the file is there. A Bazel test only sees its declared
runfiles, so the test failed with

    parse store.go: open store.go: no such file or directory

in //cmd/bd/doctor/fix:fix_test (--config=ci) and in shard 1 of
//cmd/bd/doctor/fix:fix_dolt_test (--config=doltserver), which runs the
same binary through data = [":fix_test"].

Declare store.go as data on fix_test.

* test(fix): red — the pre-enable scan must check the whole issue, as the mint does

bd config set versioned-history.enabled true runs a check over every issue
before it turns history on. The check looked at each issue's metadata alone,
while the mint canonicalizes the whole issue. A gate's timeout is a
time.Duration that a version holds as nanoseconds, so a gate whose timeout is
past 2^53-1 ns (about 104.25 days) holds a number outside the I-JSON
exact-integer range: the mint refuses every write to it, the scan never
looked, and the switch could be turned on over it.

- The agreement test's domain widens from metadata documents to whole issues:
  the metadata corpus, plus timeouts at 2500h (admitted), 2600h, 8760h and
  -2600h (refused), exactly 2^53-1 ns (admitted) and 2^53 ns (refused).
- The scan test covers a gate over 105 days and expects the verdict to name
  the field, "timeout".
- The mint-level refusal test covers the same gate and expects the refusal to
  name the field as well as the literal, for a number in metadata too.
- An end-to-end test on the direct route and one on the proxied route create
  gates with bd gate create --timeout, expect the enable to refuse and name
  them, and expect the issues the scan names to be exactly the issues whose
  next write the mint refuses.
- Three guards keep the scan's lite read from hiding a number: every column it
  leaves out is a string or a list of strings, nothing the mint adds and the
  scan does not load can hold a number, and no exported check over part of an
  issue exists beside the one the mint runs. Each guard is also run against
  shapes it must refuse, so it cannot pass by finding nothing.

The new symbols (CheckIssueVersionable, FindUnversionable and the Field of
UnversionableIssue) do not exist yet, so the issueops tests do not build.

* fix: green — the pre-enable scan runs the mint's own check over the whole issue

The check behind bd config set versioned-history.enabled true looked at each
issue's metadata alone. The mint canonicalizes the whole issue, and a gate's
timeout, a time.Duration held as nanoseconds, is a number outside the I-JSON
exact-integer range once it is past 2^53-1 ns (about 104.25 days). The mint
refuses every write to such a gate, the scan cleared the switch over it, and
the first write after that failed.

CheckIssueVersionable is now exactly canonicalDurableState's refusal over the
whole issue, and FindUnversionable runs it over every issue the mint would
version. It replaces FindUnversionableMetadata, and CheckMetadataVersionable
is gone, so no check over part of an issue is exported beside the one the mint
runs. A refusal names the top-level field it found the value in ("metadata",
"timeout") as well as the literal, on the mint and on the scan; the field is
located only after a refusal, so a write that is admitted still reads the
document once. A gate timeout above 2^53-1 ns is unrecordable by policy: it is
refused at the write, named by field and documented, not re-encoded.

The refusal at the switch gives the fix for each field. For a number in
metadata that is bd update --metadata or --unset-metadata, which works with
history on or off because a version is recorded from the state after the write.
For a gate timeout, which bd update cannot change, it is removing the gate with
bd delete, or setting the timeout_ns column with bd sql while history is off;
bd sql needs a server-backed store, and the text says so. It also says the
check runs once, now: a writer that does not record (an older bd, bd sql, a
pull from a clone that had history off) can add such a row afterwards, and the
first write to it then fails with the same refusal. Making the check and the
write of the setting one transaction would narrow that window, not close it.

Corrected in the comments, the docs and the changelog as well: an import is not
a path that skips the mint, since it goes through the same batch create; and
Dolt's JSON column collapses duplicate keys when it stores a document, so on a
Dolt store neither the scan nor the mint can see one, and the duplicate-key
refusal is a defence for a backend that keeps them.

* test(fix): red — the listing must answer as the as-of read does; the SDK opens must be accounted for

Three review findings on the versioned-history change and one wording fix,
tests first.

- A removed version's restriction. AsOfReadInTx answers a stored value outside
  the four-value vocabulary (a NULL or empty column, "live", a hyphenated or
  arbitrary string) as unknown, but ListVersionsInTx hands the column back as
  stored, for every row, and the label bd versions prints passes such a string
  through. A table test runs the listing and the as-of read on the same stored
  value, for a version that was removed and one that was not, and expects the
  same answer, and none for a version that was never removed. The label gets
  the same out-of-vocabulary values and is expected to render what unknown
  renders.
- The SDK's store opens. The shared construction guard now scans the repository
  root as well, so the Go SDK's five opens (Open, OpenFromConfig and OpenGated,
  and OpenBestAvailable in each build-tag twin) are noticed. Nothing accounts
  for them yet, so both guards fail on exactly those five sites.
- A bytes bound beside the allocation-count bound on the number admission
  gate. A classifier that builds the number it is asked about makes one
  allocation whatever the exponent, so the count bound passes it, and the bytes
  grow. The fewest bytes one call allocates for a literal with a huge exponent
  must be no more than for a small one plus a margin. It passes today, because
  the gate never builds the number; it fails for a gate that does.
- The refusal at the switch names the gate remedy as bd delete <id> --force,
  says that without --force bd only previews, and says that removing the gate
  unblocks anything it was blocking. The embedded and proxied end-to-end tests
  expect all three.

* fix: green — the listing answers as the as-of read does; the SDK opens are accounted for

Three review findings on the versioned-history change and one wording fix.

A removed version's restriction. A stored removed_restriction is a free string,
and only the four values of the vocabulary are answers. The as-of read already
reported anything else (a NULL or empty column, "live", a hyphenated or
arbitrary string) as unknown, but the listing handed the column back as stored,
for every row, and the label bd versions prints passed it through.
NormalizeRemovedRestriction is now the one place that decision is made.
AsOfReadInTx calls it; ListVersionsInTx calls it for a version that was removed
and gives a version that was not removed no restriction; and restrictionLabel
renders through it. The listing, its JSON and the label now say what the as-of
read says for the same row.

The SDK's store opens. The Go SDK's five opens (Open, OpenFromConfig and
OpenGated, and OpenBestAvailable in each build-tag twin) apply neither per-store
switch, the events journal's or versioned history's, and they still do not.
Applying them would change the library's behavior for every embedder, so that
is a separate change; this one declares the boundary. The shared construction
guard now accounts for the five with one exemption reason that is true of both
switches, so a new SDK open is noticed rather than missed, and its staleness
check keeps the exemptions honest. The docs gain one bullet saying that writers
built on the SDK do not apply the switch.

The bytes bound on the number admission gate needs no production change: the
gate never builds the number, which is what the bound pins.

The refusal at the switch now names the gate remedy as bd delete <id> --force,
says that without --force bd only previews, and says that removing the gate
also unblocks anything it was blocking. The docs line says the same. For a gate
that blocks an issue, bd delete without --force prints a preview and exits 1,
and with --force the gate is removed and what it blocked is ready again.

* test(bazel): declare the root package's Go source as data for //cmd/bd:bd_test

The shared store-construction scan now reads the repository root, where the Go
SDK's opens live, but //cmd/bd:bd_test declared none of the root package's
source. Under bazel test the scan therefore saw an empty directory, reported
the five SDK exemptions as no longer matching a store construction site, and
failed the events-journal guard, while go test, which sees the whole checkout,
passed.

List the root in tools/bazel/go_srcs.py so the generator emits its managed
go_srcs filegroup there, and add that filegroup to bd_test's data. The new
block is the generator's own output and nothing else changes.

* ci: list the switch scan's proxied gate-timeout test in the 30-shard manifest

Upstream main (#7129) moved the proxied-server tier to 30 duration-balanced
shards, and TestProxiedShardManifestGeneratorNotStale requires every
TestProxiedServer*/TestServerMode* test exactly once in that block. This
branch adds TestProxiedServerVersionedHistorySwitchRefusesAnUnrecordableGateTimeout
(cmd/bd/versioned_history_switch_scan_proxied_integration_test.go), so the
block needs its line. Written by the generator itself:
python3 scripts/ci/gen_proxied_shard_manifest.py 30 --weights=duration --write
(one line, shard 8); --check exits 0 afterwards.

Rebase note (maintainer): regenerated on current main, whose proxied-server
tier uses 34 shards; the generator's incremental output
(gen_proxied_shard_manifest.py 34 --weights=duration --write) placed this
test at shard 6 of the 34-shard block.

* ci: list the PR's embedded versioned-history tests in the embedded shard manifests

#7156 made the embedded-Dolt tier's shard manifests exact-once (50 cmd
shards, 15 storage shards), so every embedded test must be listed. This PR
adds five embedded cmd tests and three embedded storage tests; this commit is
exactly what the generators write for them:

  python3 scripts/ci/gen_embedded_cmd_shard_manifest.py 50 --weights=duration --write
  python3 scripts/ci/gen_embedded_storage_shard_manifest.py 15 --weights=duration --write

Incremental: 8 lines added, none moved.

Rebase note (maintainer): regenerated on current main, whose embedded
manifests use 100 cmd shards and 40 storage shards; the generator's
incremental output placed these tests at cmd shards 55
(TestEmbeddedVersionedHistorySwitchScansEveryRowNotOnePage), 56
(...SwitchRoundTrip), 57 (...SwitchRefusesUnversionableMetadata), 58
(...SwitchRefusesAnUnrecordableGateTimeout) and 59
(...SetManyRefusesUnversionableMetadata) of the 100-shard block, and at
storage shards 5 (TestRecordVersionKeepsSubSecondChangeAt), 6
(TestVersionedHistoryKeepsNumbersTheStoreDistinguishesApart) and 23
(TestAsOfReadContract) of the 40-shard block.

* test: exempt OpenBestAvailableWith, where #7188 moved the SDK's store construction

Upstream #7188 split beads.OpenBestAvailable: it now delegates to the new
OpenBestAvailableWith, which builds the store. This PR's two construction
guards (TestEveryStoreConstructionActivatesTheEventsJournal and
TestEveryStoreConstructionActivatesVersionedHistory) share one exemption
list keyed by function, so after merging main they flagged
OpenBestAvailableWith as an unexempted construction site and reported the
OpenBestAvailable entries as stale. The two build-tag twins' keys now name
OpenBestAvailableWith. The reason is unchanged: it is the SDK's library
surface for embedders, and no code in the bd binary calls it. gofmt
realigned the three neighbouring entries.

* chore(schema): renumber removed_restriction to 0071 after #6943's 0070

Migrations are numbered by landing order. #6943 (participation_generation,
the versioned-history write fence) lands first and owns main-plane 0070 and
ignored/0028, so this PR's removed_restriction migration moves to the next
free slot, 0071. The rename is done here, as one maintainer commit, rather
than inside the replayed commits, so the history shows exactly what changed
relative to the PR as reviewed; the intermediate commits of the rebase carry
a duplicate 0070 and are not expected to build.

- migrations/0070_add_removed_restriction.{up,down}.sql become 0071_*, with
  their header comments updated (the SQL itself is unchanged). Neither file
  exists on main or on #6943, so this is not an edit to a shipped migration.
- cli_migrations.go: the CLI-bundle override is keyed to the new file name
  and renamed cliMigration0071AddRemovedRestriction, and its const now sits
  after 0070's.
- migration_0070_add_removed_restriction_test.go becomes
  migration_0071_add_removed_restriction_test.go; its tests and messages say
  0071, and it now owns the one LatestVersion pin (71).
  TestLatestVersionIncludesMigration0070 in #6943's participation_generation
  test is superseded by it and replaced by the usual pointer comment, and
  0069's pointer comment now names the 0070 file that superseded it.
- schema_test.go hygiene comments, BUILD.bazel embedsrcs/srcs, and the
  asof_read.go comment that cited the migration number.

Checked: go test ./internal/storage/schema/... (with the dolt CLI lane,
including TestAllMigrationsSQLAppliesThroughDoltCLIAndRecordsLatestVersion)
and scripts/check-migration-hygiene.sh against origin/main and against
#6943's tip.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(versions): reconcile #6661's tests and docs with #6943's write fence

#6943's write fence (design §16.2b) makes an update-shaped write to a legacy
record -- one whose participation_generation is NULL because it was created
while versioned history was off -- a skip, not a refusal: the fence returns
before the snapshot is canonicalized, so nothing is versioned and the mint's
admission gate never runs. Several of #6661's tests seed rows with history
off and then expect a later write to them to be versioned or refused, and
several comments say every later write to an offending row fails. On a tree
that carries both PRs those tests fail and those comments are wrong.

This ports, without changing their substance, the fence-integration commits
Jim Wordelman wrote for the donnabox integration fork, where both PRs were
already folded together:

- eff898db80 / 5d2e5a2ac4 (issueops, dolt, embeddeddolt): the mocked mint
  sequence scripts the fence's participation_generation read; the number
  fidelity test splits phase 2 into 2a (a legacy row is written, not
  versioned, not refused, stays NULL) and 2b (a participating row is
  refused and left untouched); TestRecordVersionInTxSkipsALegacyRowBefore-
  TheAdmissionGate pins the fence's order against the gate.
- e0ad9d1e3c / 1528022dfa (cmd/bd): the switch round trip and the routing/env
  end-to-end tests build the rows they expect a version for after the setting
  is on, and assert that the pre-setting seeds gain nothing.
- 18b9015bf5 (cmd/bd): the embedded and proxied gate-timeout scan tests assert
  the whole scan-vs-mint truth table (legacy offenders are named by the scan
  but not refused by the mint; participating offenders are both).
- 5929ebd50c / 8ce872be03 (comments, CHANGELOG, configuration reference): the
  mint's refusal binds rows that participate in history; the pre-enable scan
  cannot see participation, so it over-approximates in the safe direction.

Left out on purpose: the fork's mintSkipsRow helper and defer-wake sweep
change, and its bd doctor "Versionable Issues" check and the docs that point
at it; none of those are part of #6661 or #6943.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Jim Wordelman <jim@wordelman.name>

* test(versions): adapt #6661's guards to what main changed under it

Three test-side adjustments for code that landed on main after #6661's last
merge of it; no production code changes.

- internal/storage/leg_contract_wiring_test.go: main gained an HTTP storage
  leg, and TestEveryLegWiresEveryRoleContract requires every leg to name or
  waive every role contract. The five R7.1 as-of read contracts this PR adds
  are a storage capability the HTTP client does not serve (internal/httpclient
  answers AsOf with an explicit unsupported error, and the wire has no
  instant/address selector or gone/unknown refusal), so they are waived on
  the http leg with that reason, as the donnabox integration fork does.
- cmd/bd/versioned_history_construction_test.go: main split
  newDoltStoreFromConfig into two thin wrappers over openDoltStoreFromConfig,
  which is where the activation defer now lives (this rebase's resolution of
  the activation commit), so the factory guard keys on that body.
- cmd/bd/versioned_history_serve_e2e_test.go: main's
  TestProxiedGatedTestsRunInTheProxiedLane requires a test that skips unless
  BEADS_TEST_PROXIED_SERVER=1 to be one the proxied lane selects, so
  TestVersionedHistoryServeActivatesTheProviderFromItsOwnRow becomes
  TestProxiedServerVersionedHistoryServeActivatesTheProviderFromItsOwnRow,
  and the 34-shard proxied manifest gains its line (written by
  gen_proxied_shard_manifest.py 34 --weights=duration --write, incremental;
  --check passes).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* chore(bazel): emit go_srcs.py's managed blocks for internal/versionedhistory

Main's tools/bazel/go_srcs.py now manages a repo_files filegroup block in
every package BUILD file, lists those filegroups in the root BUILD.bazel's
aggregates, and lists every go_library in tools/bazel's release_cross_build
target. internal/versionedhistory is a package this PR adds, so after the
rebase those managed blocks were stale and TestBazelGoSrcsBlocksCurrent
(go test ./scripts/...) failed. This is the generator's own output
(python3 tools/bazel/go_srcs.py) and nothing else; --check passes
afterwards.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(versions): don't publish change_agent in bd versions --json

`bd versions <id> --json` prints storage.IssueVersion as-is, so every JSON
tag on that struct is part of a public contract. Provenance V2
(gastownhall/beads#7510, as ruled by the maintainer) removes the ad hoc
agent field in favour of an actor chain; shipping a "change_agent" key now
would make a public contract out of a field that is on its way out.

IssueVersion.ChangeAgent is now tagged json:"-". Nothing else changes: the
issue_versions.change_agent column (migration 0067) stays, ListVersionsInTx
still reads it, change_actor and change_message are still published, and
the human listing never showed the agent. No doc or golden named the key.

TestVersionsJSONDoesNotPublishChangeAgent pins the absence with a non-empty
value, so an omitempty-style regression would still fail, and pins that
change_actor and change_message are still there.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(issueops): pin what the unfenced test-support mint does on #6943's new paths

#6943's two fixes to the shared mint body -- (a) no issues row means no
mint and no error, and (b) a create-shaped call on a new issues-table row
flagged no-history stamps participation_generation without minting -- were
written for its two kinds. This PR's shared body has a third,
mintUnfenced, behind RecordVersionAtInTx (the R7.1 as-of fixtures'
controlled-timestamp mint). The rebase resolution decided what (a) and (b)
mean for it; this pins that decision:

- (a) applies. It is the plane rule (FR-8: the table that holds the row
  decides its plane), not the fence, and the unfenced kind already honoured
  the flag-based form of the same rule (IsWisp). Without it, an unfenced
  mint for an unflagged wisps-table row would insert a version row that no
  issues row points at. The shared body's issues-table read now runs for
  all three kinds; only the update kind acts on the value it reads.
- (b) does not apply. The unfenced mint never declares participation, so a
  flagged row returns right after the snapshot, as before: no read, no
  stamp.
- The fence still does not apply: a legacy row is minted, and the closing
  UPDATE advances current_revision alone.

TestRecordVersionAtInTxKeepsThePlaneRuleButNotTheFence scripts each of
those three sequences exactly with sqlmock.
TestTestSupportMintsStayOutsideTheShapeSplit places RecordVersionAtInTx
outside #6943's create / dual-shape / update classification explicitly (it
is the seam's exempt twin, not a mutation entry point or mint site): it must
be in none of the shape lists, stay in versionExemptions, and reach neither
RecordVersionForCreateInTx (it never stamps) nor RecordVersionInTx (whose
gate and fence would stop the as-of fixtures minting history-off rows). The
row-lock guard's exemption comment for recordVersionAtInTx already names the
stamp-only UPDATE: #6943's own wording of it (all three UPDATE issues
statements) was carried into that exemption when the as-of read commit was
replayed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(storage): say exactly why the as-of read contracts stay waived on http

#6943 now runs most of its participation-generation contracts over the
served wire (internal/httpclient/served_participation_generation_test.go):
their subject is a create or an update, which the client can send, and the
fixture reads the reference store's raw columns. This checked whether the
five R7.1 as-of read contracts can be wired the same way. They cannot,
cheaply or otherwise, without a new wire operation:

- their subject is the read itself (AsOfReadFixture.Resolve), and R7.1's
  read, issueops.AsOfReadInTx, is reached by no publicops role and served
  by no wire operation;
- the wire carries neither selector (an instant or a served version
  address) nor the gone/unknown refusal the read answers with;
- the client's only point-in-time read, DoltStorage.AsOf, is a Dolt
  commit-ref read, not R7.1, and is on the client's D8 unsupported
  allowlist.

Binding Resolve to the reference store behind the served env would pass,
but as a property of that store, which the store legs already pin, never
of the client. So the waiver stays; its comment and reason string now say
the above instead of the earlier shorthand ("HTTP AsOf is explicitly
unsupported"), which conflated the client's commit-ref AsOf with R7.1.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(versions): describe what the write fence does to beads that predate recording

With #6943's write fence (design §16.2b) a bead that existed before
versioned history was turned on (participation_generation NULL) is legacy:
an update-shaped write to it is skipped, so it is never versioned however
often it is edited. Several user-facing strings from #6661 still described
the behaviour before the fence, and an independent review reproduced each
with a built bd (a legacy bead edited twice after enabling still lists no
versions; a legacy bead the switch refuses over is written without refusal
under BD_VERSIONED_HISTORY_ENABLED=1). Text only; no behaviour changes.

- bd versions help: only beads created while recording is on are versioned;
  a bead that already existed is legacy and editing it records nothing; no
  backfill. (Was: "a bead edited yesterday shows nothing until it is
  edited again".)
- bd versions empty listing: drops "yet" and says why the list is empty -- a
  bead created while recording is on always has its create version, so an
  empty list means the bead predates recording (or was made by a writer
  that was not recording), or is no-history or a wisp. The line still
  starts "No versions recorded for <id>", which the switch round-trip test
  pins.
- bd versions with recording off: only beads created after it is turned on
  are versioned.
- bd config set versioned-history.enabled true refusal: Jim's scan stays as
  conservative as he made it -- it cannot see participation, and a legacy
  row can still be minted create-shaped if it re-enters the issues plane --
  but the text now says a listed issue's writes are refused only if it
  participates (created while history was on, or entered the issues plane
  while it was on), and that legacy issues are listed because the check
  cannot tell. The closing line says the same of a value a non-recording
  writer adds later.
- docs/reference/configuration.md: the section opener says every accepted
  change to an issue created while history is on is snapshotted, and that a
  pre-existing issue is legacy; the "participates" parenthetical drops "or
  written" (a write never stamps participation; only a create-shaped mint
  does) and names entering the issues plane, matching the refusal text.

CHANGELOG already says "if it participates in history" and is unchanged.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(versions): keep updated_at across update-shaped version mints

The update-shaped mint ended with UPDATE issues SET current_revision = ?
WHERE id = ?. current_revision always changes there, so the column's
ON UPDATE CURRENT_TIMESTAMP clause fired: with versioned history on, every
label, dependency or import write moved issues.updated_at to the session's
wall clock (local time, read back as UTC). An import with
RejectStaleUpserts then stored the wall clock instead of the source's
updated_at, and the source's next genuine edit was rejected as stale. The
statement now assigns updated_at to itself, the rule #4298, blocked_state.go's
mark/unmark templates and the create-shaped mint statements already follow,
so history on leaves updated_at exactly where history off does. It is the
one statement every production update-shaped mint ends with.

Tests:
- TestEmbeddedMintLeavesUpdatedAtWhereHistoryOffDoes (embeddeddolt) ages a
  bead, adds and removes a label and a dependency, and runs two
  RejectStaleUpserts imports, with history off first as the control. With
  the fix reverted, history on fails all six arms (updated_at reads the
  wall clock, and the later source edit is rejected as stale); history off
  passes either way.
- RunParticipationGenerationFlagOffSkipsUpdateOnParticipatingRecord, wired
  on the dolt, embeddeddolt, uow and http legs: a record created with the
  flag on and updated with it off mints nothing and keeps its
  current_revision. Every other flag-off case uses a legacy record, which
  the fence stops first, so with RecordVersionInTx's gate removed this is
  the only case that fails, on all four legs. That gate is the single entry
  for every production update-shaped mint, so the one update arm covers
  label and close writes too.
- The sqlmock pin of the statement (version_history_preenable_scan_test.go)
  and the journal-completeness exemption comment follow the new SQL.
- The createbatchequiv issues_updated_at goldens of small, depadd,
  import458 and import458-reject-stale are re-recorded. They pinned the
  wall-clock bump, so each row minted a second time (eq-n2; eq-db1 and
  eq-dl; eq-b3 and eq-b250) had dropped out of the compared set; CI's
  embedded and server legs both printed the new digests. Recorded on
  embedded Dolt, the fix-reverted tree reproduces every committed golden
  byte for byte, HEAD moves only these four digests, and its rows equal
  the history-off rows in every scenario. GoldenDirEnv's comment lists it
  as the goldens' third deliberate change.
- The new embedded test has its line in the storage shard manifest
  (gen_embedded_storage_shard_manifest.py 40 --weights=duration --write),
  so TestStorageEmbeddedShardManifestGeneratorNotStale stays green. It sits
  after TestEmbeddedEveryAcceptedLifecycleMutationMintsOneVersion rather
  than at the end of the file, where #7646 appends its own test, so the two
  PRs still merge without a conflict in that file.

Docs:
- CHANGELOG: the feature's two entries under ### Fixed described fixes to a
  switch no released bd has. One ### Added entry replaces them, covering
  bd versions, the versioned-history.enabled switch and its per-store
  semantics, the pre-enable check, the single-writer caveat, migration 0071
  and the bd version <arg> refusal.
- The bd versions help, its recording-off refusal and empty-listing note,
  and the opening paragraph of the configuration doc's Versioned Issue
  History section now state the rule the pre-enable refusal prints: a bead
  participates if it was created, or entered the issues plane (a wisp
  promoted or made persistent), while recording was on.

bd version: the "did you mean bd versions" refusal now runs after the
command's metrics event is created, like the other in-RunE argument
refusals, so a mistyped bd version <id> is counted as a version invocation
instead of going uncounted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen

---------

Co-authored-by: Jim Wordelman <mayor@quad341.uk>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Julian Knutsen <julianknutsen@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-18 17:28:16 +00:00
Jim Wordelman
ce70b261c5
storage: add participation_generation (migration 0070) and the versioned-history write fence (#6943)
* schema+storage: restore 0068, move participation_generation to 0070 (be-blvmg)

MAYOR RULING 2026-09-29 on tracker be-waare: migration 0068 shipped to
origin/main as attribution_status-only (be-v33pa). be-h89oq's lineage had
appended participation_generation steps 4-5 to that same file while it was
still unmerged, which check-migration-hygiene.sh Check C correctly rejects
once 0068 is frozen on main. This restores 0068 to origin/main's exact
content and re-homes the participation_generation column (issues + wisps,
design section 16.3 steps 4-5, be-dt74u amendment) in a new migration 0070,
with its clone-local ignored/0028 twin for wisps. Ports back the
participation_generation write-fence consumer code (RecordVersionInTx in
internal/storage/issueops/version_history.go and its callers) and the
dualwrite contract tests that exercise it.

RED: the CLI-bundle direct-DDL override for 0070's guarded PREPARE/ALTER
pair is not yet registered in cli_migrations.go, so
TestBundleMigrationsWithPreparedALTERAreOverriddenOrJustified,
TestMigration0070AddsParticipationGeneration, and
TestAllMigrationsSQLUsesDirectDDLForKnownCLIIncompatibilities fail --
correctly: a fresh CLI-built bundle would silently omit the column
(dolthub/dolt#11345) while a runtime-migrated store has it.

git diff origin/main -- internal/storage/schema/migrations/0068_*.sql is
empty, confirming 0068 is untouched.

Agent-Signature: model=claude-sonnet-5; reasoning=high

* schema: register CLI-bundle direct-DDL override for migration 0070 (be-blvmg)

GREEN: adds cliMigration0070AddParticipationGeneration and its dispatch
entries (cliCompatibleMigrationSQL, cliSubstituteAssumesWispTables), the
same pre-2.3 Dolt CLI PREPARE/ALTER escape hatch 0067's wisps.current_revision
and 0068's attribution_status already use (dolthub/dolt#11345). Makes
TestBundleMigrationsWithPreparedALTERAreOverriddenOrJustified,
TestMigration0070AddsParticipationGeneration, and
TestAllMigrationsSQLUsesDirectDDLForKnownCLIIncompatibilities pass: a
fresh CLI-built bundle now carries participation_generation on both
issues and wisps, matching what a runtime-migrated store gets.

scripts/check-migration-hygiene.sh checks A-D pass; 0068 remains
byte-identical to origin/main.

Agent-Signature: model=claude-sonnet-5; reasoning=high

* build: bazel-sync BUILD.bazel for migration 0070 files (be-blvmg)

make bazel-sync (gazelle) picked up two files the RED commit added that
the hand-ported BUILD.bazel hadn't listed yet: 0070's embedded .up.sql in
the go_library embedsrcs and migration_0070_add_participation_generation_test.go
in the go_test srcs.

Agent-Signature: model=claude-sonnet-5; reasoning=high

* test(feat): red — Fix: enable history before fixtures in TestDualWriteDeleteNeighborRewriteRidesTheDeletesDoltCommit — PR #6943's §16.2b write fence leaves its flag-off neighbor legacy, so the delete mints 0 rows (stacked on 44418b3) (refs be-fyaja)

* feat: green — Fix: enable history before fixtures in TestDualWriteDeleteNeighborRewriteRidesTheDeletesDoltCommit — PR #6943's §16.2b write fence leaves its flag-off neighbor legacy, so the delete mints 0 rows (stacked on 44418b3)

(refs be-fyaja)

* test(conformance): pin that a wisp reaching the issues plane begins its history (be-31mnz)

Review feedback on #6943, blocker 1: with the flag on, a promoted wisp is
never versioned. None of the five RunParticipationGeneration* cases reaches
a promotion, so nothing caught it.

Two contract cases, wired on all three legs:

- RunParticipationGenerationWispPromotionBeginsHistory: create a wisp with
  the flag on, promote it, and require a stamped participation_generation,
  one version row, and +1 on the next update.
- RunParticipationGenerationWispPersistenceMoveBeginsHistory: the same
  check for a persistence-mode update (bd update --persistent). This mints
  at a different site than promotion on every leg: the guarded update's one
  end-of-update mint on dolt and embeddeddolt, and
  MoveIssuePersistenceInTx's own on uow.

"Wisp" is in the names because this contract already says "promoted" in
R2.3's sense (ProceedsForPromotedRecord), which is a different operation.
The fixture gains CreateWisp, PromoteWisp and MovePersistent closures, and
each leg's FixtureKitIsWired checks them.

Red at this commit on the embedded leg: after the arrival there are 0 rows
and participation_generation is NULL, and after the update there are still
0 rows. A flag-on durable create goes 1 -> 2 in the same store.

* fix(issueops): mint a wisp's first issues-plane version create-shaped (be-31mnz)

Review feedback on #6943, blocker 1. A promotion, or a persistence move out
of the wisps table, inserts a new issues row. A wisp never declares
participation (FR-8), so that row's participation_generation is NULL. Both
sites minted through the update-shaped RecordVersionInTx, and the section
16.2b fence read the NULL as legacy and skipped the mint. Nothing stamps the
column later, so every later update skipped too, and a record created with
versioned history on was never versioned.

These sites now use RecordVersionForCreateInTx, which stamps the column and
mints version 1, as a create on the issues plane does:

- PromoteFromEphemeralInTx (promote.go).
- MoveIssuePersistenceInTx's move out of the wisps table.
- ExecuteUpdate's one end-of-update mint (execution.go), when its
  persistence patch did that move. This is the site bd update --persistent
  reaches on the dolt and embeddeddolt legs, and the review did not list
  it. PersistenceMoveResult.EnteredIssuesPlane tells ExecuteUpdate which
  entry point to use.

One site the review named stays update-shaped on purpose:
MoveIssuePersistenceInTx's same-plane repair, which clears a stray wisp
flag on a row already in issues. That row keeps its own
participation_generation. A row created with the flag on still mints. A
legacy row stays legacy, because a repair is an ordinary write and R2.3
says an ordinary write does not promote it. A create-shaped mint there
would stamp the legacy row and start versioning it.

* docs(schema): 0070 must precede turning versioned history on (be-31mnz)

Review note on #6943 (schema skew). The write fence reads
participation_generation on every flag-on mint of an issues-plane row. On a
store that has not applied 0070 the read fails, and the mutation fails with
it instead of skipping. The migration header now says so, and says that the
flag-off path (the default) returns before the read, so a store below 0070
writes exactly as before.

Comment only. 0070 is not on main, so check-migration-hygiene's frozen-file
check does not apply to it, and no test pins its bytes.

* test(dolt): a persistence repair keeps the row's own participation (be-31mnz)

Pins why MoveIssuePersistenceInTx's same-plane repair stays update-shaped
while promotion and a move out of the wisps table became create-shaped (the
previous commit). The repair clears a stray wisp flag on a row that already
lives in issues, so the row keeps its own participation_generation.

TestPersistenceRepairKeepsTheRowsOwnParticipation runs the repair with
versioned history on for a legacy row (created with the flag off) and a
participating row (created with it on), through both repair entry points:
MoveIssuePersistenceInTx and a guarded update with a persistence patch. They
mint at different sites. The legacy row must stay NULL and mint nothing,
since R2.3 says an ordinary write does not promote it. The participating row
must mint one version. participation_generation must not change in either
case.

It fails if the repair mint is create-shaped, which would stamp and start
versioning the legacy row.

* test(dolt): a delete's rewrite of a legacy neighbor mints nothing (be-31mnz)

Folded in from bee's R1 on #6961, note 1. Once
TestDualWriteDeleteNeighborRewriteRidesTheDeletesDoltCommit started creating
its neighbor with history on, nothing covered a delete whose neighbor was
created with history off. Under the section 16.2b fence, the expected result
for that case is different: the rewrite still runs, but it mints nothing.

TestDualWriteDeleteSkipsALegacyNeighborsRewrite creates the fixtures with
history off, turns history on, and force-deletes the cited target. It then
checks four things: the rewrite ran (ReferencesUpdated=1), the neighbor has
no issue_versions rows, its participation_generation is still NULL, and
dolt_status is clean for issue_versions, store_epoch and issues.

* test(embeddeddolt): batch-created rows are stamped with the participation generation

A batch create writes its rows first and mints each created row's first
version last, after the creation-time edges land. The mint has to stay
create-shaped for rows the batch inserted (stamping participation_generation
from the store epoch) and update-shaped for rows it only upserted over
(skipped while the row is a legacy row), so the per-row isNew has to survive
the deferral to the batch's final mint pass.

Pin that on the embedded engine:

- brand-new explicit-id rows are stamped with the store epoch and get one
  version row each, and match the same input run one row at a time with the
  batch-create fast paths off (5 rows, and 250 rows across three INSERT
  statements);
- a legacy row the batch upserts over keeps a NULL participation_generation
  and gets no version row, while the new rows beside it are stamped;
- with versioned history off, no row is stamped and no version row is written.

* ci(bazel): enroll the participation-generation tests in the embedded shards

The embedded-Dolt shard manifest has to name every top-level Test* in
internal/storage/embeddeddolt, or the exact-once coverage check fails.
The two participation-generation fence tests were written before that
check landed on main, and the batch-create stamping test is new, so none
of the three is listed yet. Add them to the 15-shard block as the
generator's incremental output: three inserted lines, no existing
assignment moves.

Also list the new batch-create test file in the embeddeddolt_test srcs,
in sorted position, so the Bazel target compiles it alongside the Go
build. Its imports are already among that target's deps.

Rebase note (maintainer): on current main the embedded storage manifest
uses a 40-shard block; the generator's incremental output placed these
three tests in shards 18, 19 and 39 and re-sorted one existing shard-12
line.

* fix(issueops): keep a created row's updated_at when its first version is minted

Minting a version is bookkeeping. It must not change a user-visible column
of the row it versions.

The create-shaped mint stamps participation_generation with an UPDATE of the
issues row. issues.updated_at is ON UPDATE CURRENT_TIMESTAMP, so an UPDATE
that changes any other column and does not name updated_at rewrites it with
the clock. Before the stamp, a row's first mint left current_revision at its
column default of 1, so the UPDATE changed nothing and the clause never
fired. The stamp moves participation_generation from NULL to the store
epoch, so the clause fires on every row whose first version is minted and
replaces the updated_at an import supplies.

Set updated_at to itself in the stamp, as the blocked-state recompute in
issueops/blocked_state.go already does, so the timestamp the row was created
with survives. Add a regression that batch-creates rows with a supplied
updated_at and requires both the stamp and the unchanged timestamp.

* test(createbatchequiv): re-record the digests the participation-generation fence moves

The fence does not re-mint rows that predate the stamp: a row whose
participation_generation is still NULL gets no new version when a batch
touches it. Main mints one, so six recorded entries move.

The issue_versions entry of five scenarios loses the versions main mints
for rows that predate the stamp: three rows in small, all eleven in
depadd, all fifteen in waitsfor, four in import458 and two in
import458-reject-stale. In depadd, eq-db1 and eq-dl are the only rows
main mints more than once, and each of those mints moves their updated_at
off the fixture date. The fence skips them, so the two rows keep the
fixture date and join the issues_updated_at entry.

No other table, list or counter digest changes, and the fast and per-row
bodies already agree in every scenario. The digests were recorded with
the embedded backend. The doc comment on GoldenDirEnv now says which
entries were re-recorded.

* test(storage): waive the participation-generation contracts on the http leg

main's #7247 (6d21d337ad, "a served http conformance leg") registered a
fourth leg, http, after this PR's original base. The PR's seven
RunParticipationGeneration* entrypoints are neither wired nor waived there,
so TestEveryLegWiresEveryRoleContract/http fails on the rebased branch.

ParticipationGenerationFixture toggles the store's own versioned-history
write flag and reads the raw participation_generation column back: store
write-fence bookkeeping (design §16.2b), not a publicops role the HTTP
client serves. Waive the seven cases on http with a named reason, the same
waiver the donnabox fork carries (09e7ce05af), and keep the gap visible
until a supported HTTP surface is designed and tested.

go test ./internal/storage/ -run TestEveryLegWiresEveryRoleContract now
passes on dolt, embeddeddolt, http and uow.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(issueops): decide the fence's plane by the table that holds the row

recordVersionInTx chose the plane from the row's flags (IsWisp: Ephemeral
|| NoHistory) and then read participation_generation from the issues table
only. A wisps-TABLE row can carry neither flag, so it passed the IsWisp
return and the issues read failed with sql.ErrNoRows, aborting the whole
mutation whenever versioned history is on. Two producers:

  - bd mol squash clears an ephemeral root's flag in place
    (UpdateIssue {"wisp": false}, after the root's close, in the squash's
    one transaction); update.go and the uow repository's Update both mint
    after it unconditionally.
  - import's "wisp_plane" marker (WispPlaneOverride = &true, flags false)
    routes the row to wisps, and the batch create mints it create-shaped.

On origin/main the same paths do not fail but mint a stray issue_versions
row for a wisps-table row, an FR-8 leak (wisps are never versioned).

Fix: no issues row means nothing to version. ErrNoRows on the
participation_generation read now returns nil (no version, no error) for
update- and create-shaped calls alike. That read is the right signal and
IsActiveWispInTx is not: GetIssueInTx has already resolved the row, so
ErrNoRows on the issues read can only mean "not in issues", which is
exactly the question the fence has to answer before it stamps or advances
an issues row. It costs no extra query. IsActiveWispInTx asks the inverse
question ("is it in wisps?") with one more query per mint, and it folds
every error, including a transient one, into "not a wisp".

Tests come first, as three new conformance contracts with three new
ParticipationGenerationFixture hooks (ClearWispFlag, SquashWispRoot,
CreateUnflaggedWisp). They are wired on the dolt, embeddeddolt and uow legs
as subtests of each leg's TestParticipationGenerationFence (no new
top-level test, so the shard manifests do not change) and waived on http
with participationGenerationRawStoreWaiverReason:

  RunParticipationGenerationWispFlagClearMintsNothing      (create wisp,
      UpdateIssue {"wisp": false}: no error, 0 versions)
  RunParticipationGenerationWispRootSquashMintsNothing     (close +
      UpdateIssue {"wisp": false} in one transaction: no error, 0 versions)
  RunParticipationGenerationUnflaggedWispCreateMintsNothing (CreateIssues
      with WispPlaneOverride=&true, flags false: no error, 0 versions)

Red before the fix, green after, run both ways:
  embeddeddolt (BEADS_TEST_EMBEDDED_DOLT=1): all three FAIL with
    "versioned history: read participation_generation for pg-...: sql: no
    rows in result set"; all three PASS after.
  uow: all three FAIL with the same error; all three PASS after.
  dolt (BEADS_TEST_DOLT_SERVER=local): WispRootSquash and UnflaggedWispCreate
    FAIL with the same error; both PASS after. WispFlagClear passes both
    ways on this leg. The store-level wisp update (updateWisp) runs in a
    transaction versioned history never scopes, so it never reaches the
    mint. For the same reason the leg's SquashWispRoot hook turns the events
    journal on for its one transaction: that pins the wisp writes to the
    scoped regular transaction, as in a journal-enabled workspace, so that
    case does reach the seam.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(issueops): a create stamps participation even when the row is flagged no-history

A create-shaped mint (RecordVersionForCreateInTx) of a brand-new
ISSUES-TABLE row that carries NoHistory or Ephemeral returned at the IsWisp
check before the isCreate stamp, so participation_generation stayed NULL
although the row was created with versioned history on.

Import produces exactly that row: bd-r9uce keeps a promoted no-history
wisp's durable row on the issues plane with its stray no_history=true
(import_shared.go applyImportWispPlane: WispPlaneOverride = &false). When
bd update --persistent later repairs the flag, it is a same-plane normalize
whose mint is update-shaped by design (persistence.go: "A row created with
the flag on proceeds; a legacy row stays legacy"). The NULL column made the
fence read the row as legacy, so the repair and every later update were
fenced and the row was never versioned. origin/main versions it after the
repair (0, then 1, then 2 rows).

Fix: on a create-shaped call, a flagged row that the issues table holds
gets participation_generation stamped from store_epoch.epoch, the same
source the minting stamp uses, read in the same transaction. No version is
minted and updated_at is kept (SET ..., updated_at = updated_at), then the
call returns. Update-shaped calls on flagged rows still return before any
read. The table check from the previous commit now runs ahead of the IsWisp
branch for create-shaped calls, so a flagged wisps-table row (the common
wisp create) still mints and stamps nothing. That costs a create-shaped
call on a flagged wisp one primary-key miss on issues, and only while
versioned history is on. The store_epoch read/seed moves into
readStoreEpochInTx so both stamps share it.

Tests first:
  - New conformance contract
    RunParticipationGenerationNoHistoryIssueCreateDeclaresParticipation and
    hook CreateNoHistoryIssue. Wired on dolt, embeddeddolt and uow as a
    TestParticipationGenerationFence subtest and waived on http. With
    history on: create {NoHistory: true, WispPlaneOverride: &false} gives
    participation_generation non-NULL and 0 versions. The persistent repair
    (MovePersistent) then gives 1 version, and an update-shaped edit
    (MutateExisting, a title edit) gives 2.
  - New subtest
    TestCreateBatchParticipationGeneration/NoHistoryRowIsStampedUnversionedAndKeepsTheCallersUpdatedAt
    (embeddeddolt): a batch of such rows with a supplied updated_at comes
    out stamped, unversioned, and with updated_at unchanged.

Red before the fix, green after, run both ways:
  dolt (BEADS_TEST_DOLT_SERVER=local), embeddeddolt
  (BEADS_TEST_EMBEDDED_DOLT=1), uow: the contract FAILS on all three legs
    ("participation_generation for pg-no-history-create is NULL ...", then
    "= 0 after the persistent repair, want 1", then "went from 0 to 0 ...
    want exactly +1"). It PASSES on all three after.
  embeddeddolt subtest: FAILS before the fix ("store_epoch has no row after
    the batch, so nothing stamped the 5 rows"). With the fix minus the
    updated_at = updated_at clause it FAILS on every row's updated_at. It
    PASSES with the full fix.
No new top-level Test*, so the shard manifests do not change. The
createbatchequiv goldens are unchanged: the dolt and embeddeddolt
equivalence tests pass.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs: fix comments the rebase left stale

Comment-only.

- createbatchequiv.GoldenDirEnv said "One deliberate behavior change since
  ... nothing else moved" and then listed a second re-record, the
  participation-generation fence's. It now says there are two deliberate
  changes, puts each in its own paragraph, and scopes "nothing else moved"
  to apply.json, the only golden the first change touched.
- issueops CreateIssueResult.isNew and deferredVersion still described one
  write path. After main's #7205, a row written by deferredCreates'
  multi-row INSERT reaches finishCreateIssueInTx(..., true) directly
  (create_batch_cache.go flush) and never passes through InsertIssueIfNew
  or CreateIssueInTxWithResult. isNew is InsertIssueIfNew's result on the
  per-row path and always true for a multi-row-INSERT row. deferredVersion
  is filled from both paths.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(issueops): pin the create/update shape split in the version guard

TestEveryVersionedEntryPointMints counts RecordVersionInTx and
RecordVersionForCreateInTx as one seam. A create-shaped site that fell back
to the update-shaped entry therefore stayed green, though the fence then
reads the new row's NULL participation_generation as legacy and the record
is never versioned. The reverse case also stayed green: an update path that
reached the create-shaped entry would stamp an existing legacy row and
promote it, which R2.3 forbids.

Add the split to the structural guard. It is two small tests over the same
journalscan fixpoint and mintVersion edge discipline, with no AST changes:

- TestCreateShapedEntryPointsMintCreateShaped: the create-shaped mint sites
  (finishCreateIssueInTx, CreateIssuesInTxWithContext,
  PromoteFromEphemeralInTx, moveIssuePersistenceInTx, executeUpdate) must
  each call RecordVersionForCreateInTx themselves. A delegate is not enough
  here, because executeUpdate runs its move with mintVersion=false and a
  batch switches finishCreateIssueInTx's mint off at run time. Every
  create-shaped entry point, the delegates and #7205's
  deferredCreates.flush included, must reach it.
- TestUpdateShapedEntryPointsNeverMintCreateShaped: every other
  versionedEntryPoint must not reach RecordVersionForCreateInTx. The
  exception is ApplyBatchInTx, which runs create and update items together
  and is listed as mixed with a reason.

Mutation-checked, each probe reverted afterwards. executeUpdate's
EnteredIssuesPlane mint, PromoteFromEphemeralInTx's mint, or the batch's
deferred create-shaped mint switched to RecordVersionInTx: each fails the
create half. update.go's end-of-update mint switched to
RecordVersionForCreateInTx: fails the update half for UpdateIssueInTx and
every entry point that writes through it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(issueops): hold the dual-shape mint sites to both halves of the split

f006c871c5 listed executeUpdate, moveIssuePersistenceInTx,
finishCreateIssueInTx and CreateIssuesInTxWithContext (and the
ExecuteUpdate/MoveIssuePersistenceInTx/create wrappers that delegate to
them) as create-shaped, so TestUpdateShapedEntryPointsNeverMintCreateShaped
skipped them. Each one also mints update-shaped, on its existing-row or
same-plane branch, which is the common path: a guarded update of an
existing row, the bd update --persistent repair of an issues-plane row's
flags, a create or batch create that resolved to a row already there.
Switching that branch to RecordVersionForCreateInTx stamped a legacy row
out of legacy (R2.3), and every guard stayed green.

Split the lists by what each function actually mints:

- versionCreateShapedMintSites keeps only the pure create-shaped site,
  PromoteFromEphemeralInTx. It must call RecordVersionForCreateInTx and
  must not call RecordVersionInTx.
- versionDualShapeMintSites names the four, each with its two branches.
  TestDualShapeMintSitesCallBothEntries requires each to call BOTH entries
  directly. A delegate does not count, for the same mintVersion=false and
  DeferVersionMint reasons as before.
- versionCreateShapedEntryPoints keeps PromoteFromEphemeralInTx and
  deferredCreates.flush, which passes isNew=true for every row.
- versionDualShapeEntryPoints replaces versionMixedShapeEntryPoints. It
  holds ApplyBatchInTx and the create, persistence-move and guarded-update
  entries that delegate to a dual-shape site, each with a reason. Each must
  reach both entries transitively.

The update half skips the create-shaped and dual-shape entry points, as
before. The update-shaped branch inside a dual-shape site is now the new
test's to pin.

Mutation-checked in a throwaway worktree. Switching the update-shaped mint
to RecordVersionForCreateInTx at execution.go:301, persistence.go:97,
create.go:260 or create.go:499 now fails
TestDualShapeMintSitesCallBothEntries for that site. Before this commit,
the whole issueops package stayed green under each of these switches.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(httpclient): run the participation-generation fence over the served wire

edd3b72a1a waived all eleven ParticipationGeneration contracts on the http
leg, saying the client had "no write flag to toggle and no raw column to
read". That reason was wrong. served_dualwrite_history_test.go already
flips the reference store's VersionedHistoryConfigurer and reads its raw
columns, with the client driving the mutation. The waiver also hid the
one harness that runs the fence behind the update a connected workspace
actually sends. Every store leg's MutateExisting is store.UpdateIssue,
but the served Lifecycle.Update reaches the guarded update,
ExecuteUpdate.

Add served_participation_generation_test.go on the same pattern:

- Bound to the client: Mutate is Lifecycle.Create, MutateExisting is
  Lifecycle.Update, MovePersistent is the persistence-mode
  Lifecycle.Update that bd update --persistent sends, and PromoteWisp is
  issueops.Promote over the client's Reader/Lifecycle/Commenter, the
  entry bd promote dials over http.
- Bound to the reference store, as before: SetFlag, the CreateWisp seed,
  and the raw reads.

TestServedParticipationGenerationFence runs seven cases:
CreateWithFlagOnStampsAValue, CreateWithFlagOffLeavesLegacyNull,
SkipsUpdateOnLegacyRecord, ProceedsForPromotedRecord,
FenceIgnoresSchemaSkewHatch, WispPromotionBeginsHistory and
WispPersistenceMoveBeginsHistory. All seven pass, and no case turned up a
server-side fence bug.

Four cases stay waived, now with a reason that names the actual wire gap
(participationGenerationWireGapWaiverReason):

- WispFlagClearMintsNothing: the in-place {"wisp": false} is an untyped
  UpdateIssue, and IssuePatch.Persistence moves the row instead.
- WispRootSquashMintsNothing: it needs RunInTransaction. That and
  UpdateIssue are both on the client's unsupported allowlist (D8).
- UnflaggedWispCreateMintsNothing and
  NoHistoryIssueCreateDeclaresParticipation: both need WispPlaneOverride,
  which every create shape on the wire drops
  (roleIgnoredCreateIssueMembers).

The new file is listed in httpclient_test's srcs. The served lane deals
unpinned tests round-robin, so no shard manifest names it.

Mutation-checked in a throwaway worktree. With execution.go:301 (the
guarded update's update-shaped mint) switched to
RecordVersionForCreateInTx, SkipsUpdateOnLegacyRecord and
FenceIgnoresSchemaSkewHatch fail here ("version row count for
hpg-legacy-skip went from 0 to 1"). The embeddeddolt and uow
Participation|DualWrite|Version suites stay green under the same probe.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(conformance): an update never stamps a legacy no-history row

7c00d65fa6 made the seam write participation_generation on a flagged
(NoHistory or Ephemeral) issues-table row when the call is create-shaped.
Only the early `if IsWisp(issue) && !isCreate { return nil }` keeps an
update-shaped call out of that branch, and nothing pinned it. The new
contract and its subtest covered only the positive, create-shaped side.
With the early return dropped, an ordinary update stamped a legacy
no-history row out of legacy (R2.3), and every suite stayed green.

Add RunParticipationGenerationLegacyNoHistoryIssueStaysLegacy beside
NoHistoryIssueCreateDeclaresParticipation as its negative twin:

1. Create the no-history issues-table row with the flag OFF.
2. Turn the flag on and run MutateExisting. participation_generation must
   stay NULL, with 0 versions and current_revision unchanged.
3. Run MovePersistent, the bd update --persistent repair that clears the
   flag in place. The repair is update-shaped too, so the same three
   things must still hold.

Wired on all four legs: dolt, embeddeddolt, uow and http. On http the
case exercises the client's Lifecycle.Update and persistence update.
Their precondition, the flag-off no-history create, is seeded through
the reference store, because no client create can land that row in the
issues table. That gap is also why the case whose SUBJECT is that create
stays waived on http.

Mutation-checked in a throwaway worktree, with the early return made
`if false && IsWisp(issue) && !isCreate`. The new case fails on all four
legs and is the only case that fails:
"participation_generation for pg-no-history-legacy = 1 after an
update-shaped mutation with the flag on, want NULL" (dolt server leg,
embeddeddolt, uow and http). At tip it passes on all four.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(issueops): say which table and which mint the fence comments mean

Comment and doc corrections. No code changes.

- RecordVersionForCreateInTx's doc comment covered a flagged issues-table
  row but not a wisps-table row. A row the wisps table holds mints and
  stamps nothing, whatever its flags say: there is no issues row to stamp
  (FR-8). That holds for an ordinary wisp create and for import's
  unflagged "wisp_plane" row.
- The no-history stamp branch said the bd update --persistent repair mints
  update-shaped "(MoveIssuePersistenceInTx)". That is true only on the
  unit-of-work path. On dolt and embeddeddolt the repair goes through the
  guarded update: executeUpdate runs moveIssuePersistenceInTx with
  mintVersion=false and mints once at its own end. The fixture's
  MovePersistent doc already says this.
- VERSIONED-BEADS-WRITE-PATHS.md row 13 still credited the whole wisp
  exclusion to `RecordVersionInTx returns on IsWisp` and called every case
  a NO-OP. After ec52d29030, the table holding the row decides its plane:
  only an update-shaped call on a flagged wisp returns on IsWisp. Every
  other wisps-table row, including an unflagged one, is excluded by the
  issues-table participation_generation read, which finds no row. After
  7c00d65fa6, a create-shaped call on a flagged issues-table row is
  stamped but not versioned, so the row says that too.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(conformance): the wisps-table cases check the row is still a wisp

runWispTableRowMintsNothing asserted 0 versions but never checked its own
precondition, that the wisps table still holds the row. Suppose a step
moved the row onto the issues plane through an update-shaped mint. That
mint would hit the new row's NULL participation_generation and be fenced,
so the count stayed 0 and the case passed without saying anything about
a wisps-table row.

- Add an InWispsTable hook to ParticipationGenerationFixture and bind it
  on dolt, embeddeddolt and uow as a raw wisps-table count. The
  wisps-table body now Fatals if the row is not in the wisps table after
  its steps. Each leg's KitIsWired test checks the hook. On http it
  stays nil with the other wisps-table hooks, whose cases are waived
  there.
- The dolt ClearWispFlag hook now bites. A bare store.UpdateIssue on a
  wisp goes through updateWisp, whose transaction versioned history never
  scopes. So WispFlagClearMintsNothing passed on dolt with or without the
  ec52d29030 fix, as that commit's message admits. The hook now runs the
  lone flag clear in a RunInTransaction with the events journal on: the
  journal-enabled shape the SquashWispRoot hook already uses.
- Both hooks go through a new withEventsJournal helper. It restores the
  store's previous journal setting rather than unconditionally setting it
  to false afterwards, which assumed the store started with the journal
  off.

Mutation-checked in throwaway worktrees:
- With version_history.go at ec52d29030^, dolt's WispFlagClearMintsNothing
  now fails ("read participation_generation for pg-flag-clear: sql: no
  rows in result set"). It used to pass there.
- Embedded ClearWispFlag rerouted as a flag-off persistence move followed
  by a flag-on title update: the case fails on the new precondition ("is
  not in the wisps table after flag-clear"). With the precondition
  disabled, the same probe passes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(storage): review fixups for the participation_generation fence

The review of d17b716115 approved the PR with no required changes and
raised four minors and three nits. Each one below says what changed and
why, or why it did not.

Minor 1, the upsert create's new-or-existing choice was untested. The
unit-of-work create that is not CreateOnly (bd batch, bd todo, tracker
sync and mol port) inserts through an upsert, and db.Insert's existed
probe picks the mint's shape; forcing it either way shipped green. The
new uow TestParticipationGenerationUpsertCreate pins all three outcomes:
a new row is stamped and begins its history, a legacy row the upsert
resolves to stays legacy with no version, and a participating row mints
exactly one more. No served-route case is added because every served
create is CreateOnly (uow createParams), so the http wire cannot reach
this branch. The cmd/bd proxied-server batch and todo creates call this
same use-case method.

Minor 2, an update-shaped mint still lets issues.updated_at move
through ON UPDATE CURRENT_TIMESTAMP. Not changed here. The maintainer
listed it on the PR as follow-up (2), and the approving review of
d17b716115 accepted carrying it as a follow-up. It does not change
the fence's contract as pinned.

Minor 3, the depadd and waitsfor goldens had been re-recorded to zero
issue_versions rows. Their seeds run with history off, so every seeded
row was legacy and the harness stopped checking the dependency-add,
close and reopen mints. Both scenarios now stamp their seeded
issues-table rows after the seed (stampParticipation; wisps stay
unstamped). Their goldens are back to main's pre-fence files byte for
byte: depadd mints 11 version rows and waitsfor 15. The fenced legacy
outcome stays pinned by small, import458 and import458-reject-stale. A
run scenario that mints no version row now fails rather than recording
an empty digest.

Minor 4, PromoteFromEphemeralInTx always minted create-shaped. When the
issues plane already held the id (the dual presence GH#4455 records),
its upsert resolved to that row and the create-shaped mint stamped a
legacy record (R2.3). It now keeps InsertIssueIfNew's isNew and mints
update-shaped for an existing row, as finishCreateIssueInTx does. The
version guard lists it as a dual-shape mint site and entry point, and
PromoteOverDualPresenceKeepsLegacy pins the behavior at run time.

Nit 1: with versioned history off, db.Insert skips the existed probe,
since both mint entries return before reading anything. A create costs
the round trips it did before the fence.

Nit 2: the MINT COST comment now counts the fence read (about 6 to 7
round trips) and no longer says the snapshot reads current_revision;
the fence read can select it for #6379's ordinal seed. Moving that read
now would only add a column nothing consumes yet.

Nit 3: the http leg's participation-generation wire-gap waiver opens as
PENDING and names the wire mutation that would retire each case.

Validation, in a scrubbed env with -tags=gms_pure_go on embedded and
server Dolt (all pass, none skipped):
- uow: TestParticipationGenerationFence, its FixtureKitIsWired check
  and TestParticipationGenerationUpsertCreate with its 4 subtests.
- embeddeddolt: the create-batch equivalence parts (small/depadd,
  apply/waitsfor), the parts-coverage guard, the large scenarios, and
  the fence and kit.
- dolt: TestCreateBatchFastPathsMatchPerRow_Dolt, the fence,
  FenceIgnoresSchemaSkewHatch and the kit.
- httpclient: TestServedParticipationGenerationFence.
- TestEveryLegWiresEveryRoleContract, the issueops structural guards,
  and the issueops, storage and domain/db packages.

Eleven mutants, graded in 13 runs, each fail the test meant to catch
them:
- upsert always create-shaped or always update-shaped;
- the probe guard inverted;
- promote always create-shaped or always update-shaped;
- the seed left unstamped;
- the dependency-add, close or reopen mint dropped;
- two on-disk single-entry promotes.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(storage): a wisp upsert must not stamp a legacy issues twin

domain/db's Insert chose the mint shape of its upsert (the create that
is not CreateOnly) from an existence probe of the target table. For a
wisps-table insert that probe asks the wrong plane: the seam snapshots
the issues table first (GetIssueInTx), so when the issues plane already
held the id as a legacy row, the wisp create took the create-shaped
mint and stamped and versioned that untouched twin. An ordinary wisp
write promoted a legacy record, which R2.3 forbids. bd mol port and the
proxied mol and wisp writers call CreateWisp this way. It is latent
until versioned history is enabled (#6661).

A wisps-table insert never creates an issues row, so it now mints
update-shaped, without the probe. The fence leaves a legacy twin alone,
a wisp-only id still mints nothing through either entry (FR-8), and a
participating twin mints the one version main already minted for it.
Insert's comment now records that last case: the dual presence
predates the fence, and this upsert does not refuse it the way
issueops' create path does.

TestParticipationGenerationUpsertCreate gains
WispOverLegacyIssuesRowKeepsItLegacy, which fails against the previous
dispatch (twin stamped, one version row). Also from review: the version
shape guard names the dual-shape site its parse cannot see, and the
row-lock exemption comment counts all three of the seam's UPDATE issues
statements.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Julian Knutsen <julianknutsen@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-18 13:10:57 +00:00
Julian Knutsen
bdd16d6383
fix(sqlbuild): bind time bounds as their UTC instant (#7655)
* fix(sqlbuild): bind time bounds as their UTC instant

BuildIssueFilterClauses bound every time bound (created/updated/closed/
started/defer/due, after and before) as tc.v.Format(time.RFC3339), which
keeps the value's offset. The columns hold UTC wall clock, and an engine
whose timestamp column carries no zone reads a literal's wall clock and
drops the offset, so a bound spelled +02:00 would compare two hours late.
Bind tc.v.UTC().Format(time.RFC3339): the same instant, spelled in UTC, at
whole seconds. Dolt already converts an offset literal to UTC and bd's
client resolves bounds to UTC, so no answer changes on any backend here;
this keeps the shared builder correct for every engine.

TestTimeBoundsBindTheUTCInstant pins all twelve bounds to the UTC spelling
of a +02:00 instant with a sub-second remainder.

Also correct the AfterCreatedAt comment, which justified the time.Time
cursor binding by a SQLite backend that no longer exists.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(sqlbuild): bind the keyset cursor as its UTC instant too

Review fixups on #7655.

BuildIssueFilterClauses bound IssueFilter.AfterCreatedAt as the caller's
time.Time, unconverted. The embedded Dolt driver sends a time.Time as its
wall clock with no zone (vitess BuildBindVariable formats it as
"2006-01-02 15:04:05.000000"), and the columns hold UTC wall clock, so a
cursor in another zone paged from the wrong position there: the keyset
walk with +02:00 cursors re-delivers rows on embedded Dolt.
go-sql-driver/mysql converts to the connection's zone (UTC) first, so
server Dolt was unaffected. Bind filter.AfterCreatedAt.UTC(), still a
time.Time, and reword the comment, which claimed the driver carries the
instant. TestKeysetCursorBindsTheUTCInstant pins the cursor args of both
orders.

The httpapi query.timestamp comment and the
TestListFilterTimeBoundsAreNormalizedToUTC doc justified the route's UTC
normalization by the old sink behavior (formatting a bound in its own
zone). They now say sqlbuild binds every range bound in UTC and the
route's normalization is its own guard in front of that.

The CHANGELOG bullet gets a bold lead, covers the cursor, and drops the
blank line that made the Fixed list loose.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-17 16:19:50 +00:00
Steve Yegge
46aa59cc65
test(cmd/bd): stop tests from mutating the host checkout's .git; guard it in TestMain (wy-qlvlmf) (#7650)
* wy-qlvlmf: test(cmd/bd): stop tests from mutating the host checkout's .git; guard it in TestMain

On 2026-10-16 a 'go test ./cmd/bd' run from a linked worktree rewrote the
host rig's shared .git/config (core.bare=true, extensions.worktreeConfig,
remote.origin.url replaced by metadata_test's fixture URL, user identity,
core.hooksPath) and its hooks dir, leaving every worktree of that clone
answering 'fatal: this operation must be run in a work tree' (r880).

Root causes, all one class — fixture git commands routed at the repository
the test process itself lives in instead of a fixture the test owns:
- TestRunApplyAllOK ran runApply(false) (a LIVE apply: hook reinstall,
  hooksPath mkdir) with no fixture, against the package cwd's repo.
- runGitForBootstrapTest treated dir=="" as 'inherit package cwd'.
- doctor/fix setupGitRepoInDir swallowed git config failures non-fatally
  and ran with whatever git routing env (GIT_DIR, GIT_WORK_TREE) leaked in.

Fixes:
- TestRunApplyAllOK / TestRunApplyDryRun now chdir into an isolated
  git+.beads fixture (isolatedApplyFixture + runInDir).
- runGitForBootstrapTest: empty dir gets a neutral t.TempDir() cwd, never
  the package cwd; env scrubbed of git routing vars.
- initGitRepoAt / doctor setupGitRepoInDir / doctor-fix setupGitRepoInDir:
  refuse an empty dir, fail on ANY git error, scrub routing env.
- testMainInner scrubs git routing env process-wide at startup.
- TestMain's repo guard now also snapshots the checkout's own git state
  (common .git/config, config.worktree, common hooks dir) and reddens the
  run on any diff, naming the mutated file.

Guard vacuity check: a probe test writing one key into the checkout's
.git/config passes individually and reddens the suite via the guard.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* build(bazel): auto-sync BUILD files

Applied from the bazel-sync-patch artifact of https://github.com/gastownhall/beads/actions/runs/39013861275
(`make bazel-sync`: gazelle, tools/bazel/go_srcs.py, bazel mod tidy).
Only BUILD.bazel, MODULE.bazel and MODULE.bazel.lock are ever pushed; see
scripts/bazel-autofix-push.sh.

* test(cmd/bd): scrub doctor suites, keep git fences, key-level config guard (wy-qlvlmf)

Maintainer fixup on top of #7650 by cat, for the first review's three
majors, minor and nit, and three of the second review's minors.

- Doctor suites: the leaked-GIT_DIR vector was still open in
  cmd/bd/doctor and cmd/bd/doctor/fix. Their four TestMains now unset
  the repository-location env first (testutil.UnsetGitRepoLocationEnv),
  and the three initGitTemplate copies plus
  testutil.ForceRepoLocalHooksPath run git with a filtered env. With
  GIT_DIR/GIT_WORK_TREE leaked at a decoy repo, the PR head wrote
  core.hooksPath into the decoy and failed 120 doctor tests; with this
  commit all three packages pass and the decoy stays byte-identical.

- Narrower scrub: the process-wide scrub also removed
  GIT_CEILING_DIRECTORIES and the GIT_CONFIG_* fences that
  tools/bazel/test_env.sh and scripts/ci/lib/test-env.sh set on purpose.
  testutil.WithoutGitRepoLocationEnv removes only GIT_DIR, GIT_WORK_TREE,
  GIT_COMMON_DIR, GIT_INDEX_FILE, GIT_OBJECT_DIRECTORY,
  GIT_ALTERNATE_OBJECT_DIRECTORIES, GIT_NAMESPACE and the legacy
  GIT_CONFIG (which sends `git config` reads and writes to the file it
  names, even under -C). Keys match exactly, so GIT_CEILING_DIRECTORIES,
  GIT_CONFIG_GLOBAL and GIT_CONFIG_NOSYSTEM stay. Every helper that
  built its own scrubbed env now uses it.

- Guard false reds: the TestMain guard compared the shared common config
  byte for byte, so a sibling worktree's routine branch write failed a
  passing run. It now diffs `git config --list -z` key by key.
  branch.<name>.* keys, and a remote other than origin added or removed
  whole, print a NOTE and leave the exit code alone. Any other key
  (including origin appearing or disappearing, and a changed key of a
  remote that exists before and after), any config.worktree change, any
  hook file change or an unparseable config still fails the run.

- config.worktree: the guard also snapshots a main checkout's own
  config.worktree, which sits in the common dir. With
  extensions.worktreeConfig set its keys override the shared config, and
  before this an escaped `git config --worktree` write there went
  unreported.

- Report: changed keys are named with values redacted, deleted and new
  files are reported as such, pre-run copies of the changed files are
  saved (0700 dir, 0600 files) and their path printed, and a failure to
  locate the checkout's git dirs prints a warning instead of disarming
  the guard silently.

- doctor setupGitRepoInDir stat-checks its dir like the doctor/fix copy.

New tests: internal/testutil/git_env_test.go, and
cmd/bd/test_repo_git_guard_test.go (real-git fixtures for locating the
git dirs, classifying writes, config parsing and the report).

Validation: ./scripts/test.sh ./cmd/bd in a scratch clone with GIT_DIR
leaked at a decoy (ok; decoy and the clone's .git unchanged; guard
quiet); bazel test --config=ci for testutil, doctor, doctor/fix,
fmt_test and the new bd_test guard tests; make ci-pr-lint;
tools/bazel/go_srcs.py --check.

Second review: leaving GIT_CONFIG out of the scrub fails
TestWithoutGitRepoLocationEnvKeepsWrapperGitControls and
TestForceRepoLocalHooksPathIgnoresLeakedGitDir; the old main-checkout
config.worktree path fails TestLocateRepoGitDirs/main_checkout and the
new config.worktree case; treating origin like any other remote fails
the new origin added and removed cases. In a scratch main-checkout clone
with extensions.worktreeConfig set, a test that wrote
`git config --worktree user.email` into the checkout now fails the
suite (exit 1, an ERROR naming .git/config.worktree). bazel test
--config=ci passes for testutil, fmt_test and the bd_test guard tests.

Not in this commit: the same scrub in beads_test_env_enter
(scripts/ci/lib/test-env.sh) and tools/bazel/test_env.sh, which would
cover the packages outside cmd/bd that still write to the clone under a
hook-leaked GIT_DIR. That predates this PR and is left to a follow-up.

Co-authored-by: cat <cino@ghosttrack.com>
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: cat <cino@ghosttrack.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-17 12:36:48 +00:00
Julian Knutsen
4450fe99bb
feat(http): bd blocked over bd serve (blocked.list), plus bd blocked --flags [--raw] (#7649)
* feat(http): bd blocked over bd serve (blocked.list) through issueops.BlockedLister

`bd blocked` refused against an HTTP workspace: it called the legacy
GetBlockedIssues, which the HTTP backend stubs as unsupported. It now runs
one role on every route, and bd serve publishes that role.

- Role: issueops.BlockedLister.ListBlocked(BlockedRequest{ParentID, Labels,
  LabelsAny, ExcludeLabels}) -> BlockedResult{Items}. One body,
  internal/storage/issueops.ExecuteListBlockedInTx over the unchanged
  GetBlockedIssuesInTx; labels normalized in the role (BlockedFilter);
  Items never nil and in a total order (priority, created_at desc, id).
- Legs: DoltStore (read tx), EmbeddedDoltStore (read conn), uow (read unit
  of work through the use case's GetBlockedIssues, so the
  external-dependency uow decorator applies). Census: storage.DoltStorage
  gains BlockedLister(); HookFiringStore passes it through (a read),
  telemetry wraps it, uow.notifyingProvider, externaldeps.uowProvider and
  httpapi.timedProvider build it.
- externaldeps.Store.BlockedLister is built over its own GetBlockedIssues,
  so local `bd blocked` output is unchanged. Over a remote store it takes
  the server's listing when the server advertises
  policy.external_dependencies, else decorates the server's listing here
  through EdgeReader and BatchGetter (decorateBlocked, factored out of
  GetBlockedIssues).
- Wire: GET /v0/beads/blocked (listBlockedIssues; parent, label,
  label_any, exclude_label) answering BlockedPage {items, has_more} with
  the new x-go-type-pinned BlockedIssue schema, behind the new
  blocked.list token. apigen and the wire-shape golden regenerated; the
  golden diff is additive only, so wire_revision stays 2.
  Config.BlockedLister is a required role, wired in bd serve.
- Client: httpclient.Store.BlockedLister dials through Store.dispatch, so
  an older server is refused BEFORE the dial with *wire.CapabilityError
  naming blocked.list. The legacy Store.GetBlockedIssues rides the role
  (page and row cap applied to its answer) and leaves the unsupported
  shell. Divergence ledger row F-blocked retired.
- CLI: `bd blocked` takes BlockedLister() from the store on the direct
  routes and from the provider's BlockedListerSource on the proxied route;
  runBlockedProxiedServer is deleted. --json output is the same rows in
  the same order, [] when empty.

Tests: backend/conformance RunBlockedLister (10 cases: direct blocker,
child attributed to its blocked parent, closed blocker, wisp plane, pinned
status, parent filter, label filters with normalization, order, empty
slice, request not mutated) on dolt, embedded, uow and the served HTTP
client; httpapi handler tests (mapping, unknown/repeated params,
projection, capability); httpclient stub tests (pre-dial refusal naming
the token, encoding, projection, legacy door paging); skew and vocabulary
counts; TestE2E_BlockedOverHTTP (served `bd blocked --json` equals the
local command's after mapping ids and timestamps, member order included,
external blocker named, older-server refusal with no dial); corpus
gc-blocked re-marked and gc-blocked-parent added, both pending until the
reference server advertises blocked.list.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(blocked): pin server-side external decoration, refuse has_more, serve the legacy door over http

Review fixes for the blocked.list slice.

- Server-side external decoration is now pinned on the compositions bd
  serve answers from. TestWrapUOWProviderBlockedListerNamesExternalBlockers
  lists through the wrapped unit-of-work provider's own BlockedLister
  accessor (the bd serve provider path) and requires an external-only row
  and a mixed row to carry their external: refs; building that lister over
  the inner provider fails it. TestServedBlockedListerCarriesExternalRefs-
  WhenTheServerEnforces serves the store-shaped composition
  (externaldeps.Store.BlockedLister, policy advertised) and requires the
  operation's own answer, through the undecorated client role, to carry the
  refs, and the client's decorated chain to return it unchanged; serving
  the inner store's lister fails it.
- The client refuses an answer reporting has_more with the new typed
  httpclient.TruncatedListingError instead of returning a short listing:
  the operation has no cursor, and a caller reads a missing row as
  unblocked. The legacy door inherits it.
- externaldeps.Store.GetBlockedIssues over an http store reads the extra
  sources through the served roles (EdgeReader, BatchGetter) instead of the
  raw GetIssuesByIDs / GetDependencyRecordsForIssues the http backend
  refuses, and the decoration skips the source read when there is nothing
  to load. The legacy door on the CLI chain (bd mol stale, the wisp
  commands) now answers over http in both policy modes;
  TestServedLegacyGetBlockedIssuesThroughTheDecorator pins it, paging
  included. bd ready --explain still refuses on its other raw reads; the
  CHANGELOG says so.
- Docs: the spec and the retired F-blocked ledger row state that the
  listing is unpaged, bounded by the client's 32 MiB response limit
  (refused, never read in part), and that has_more is refused.
- Contract: new cases for waits-for and conditional-blocks blockers, no
  type exclusions (gate, epic), and a stale flag with no nameable blocker
  (on the legs that can write the column: dolt, embedded; skipped
  elsewhere with the reason). Order seeds explicit created_at values with
  five rows tied on priority and created_at, two of them wisps, so only the
  id tiebreak orders them; dropping the tiebreak now fails every run.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(blocked): bd blocked --flags, the blocked flag per id from stdin

`bd blocked --flags [--json]` reads ids from stdin, one per line (blank
lines and repeats ignored), and prints the blocked flag bd holds for each:
with --json, [{"id","is_blocked"}] in the order the ids were first given,
omitting an id that names no row on either plane, [] for no input. It is
the readiness column a consumer otherwise had to read with `bd sql`, so no
such consumer needs a local-only command.

It is the FLAG, not the listing: `bd blocked` narrows the stored flag to
rows it can name a blocker for, and --flags does not, so a row whose flag
is set with no nameable blocker reads true here and is absent from the
listing. Where the external-dependency policy applies, an unsatisfied
external: blocker sets the flag, as it keeps the row out of `bd ready`.

One library entry on every route: issueops.BlockedFlags over the route's
BlockingAnnotator (the store's accessor on dolt, embedded and HTTP, the
provider's on the proxied route). Over HTTP that is listBlockingAnnotations
with is_blocked behind dependencies.blocking.isBlocked; a server without
that token is refused before the dial, naming it. No wire change. The
listing's filter flags are refused with --flags rather than ignored.

Tests: TestE2E_BlockedFlagsOverHTTP (served JSON equals the local command's
after mapping ids to titles, unknown id omitted, repeat collapsed, [] on
empty stdin, filters refused; a stale-flagged row with no blocker reads
true and is absent from `bd blocked --json`, which fails if --flags is
derived from the listing; an older server's handshake refused naming the
token with no dial); TestProxiedServerBlocked/flags_reads_ids_from_stdin on
the proxied route; corpus gc-blocked-flags, pending until the reference
server advertises the token. CHANGELOG entry.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(blocked): bd blocked --flags --raw, the stored is_blocked column

`bd blocked --flags --raw [--json]` answers the stored is_blocked column
for each id read from stdin, exactly what reading is_blocked from the issues
and wisps tables returns, without the external-dependency policy: a row
blocked only by an unsatisfied external: ref reads false (the policy-aware
--flags reads true), and a stale flag reads true either way. Same output
shape as --flags. This lets a consumer move off a `bd sql` read of the
column with no change in what it reads. The default --flags is unchanged.

The role is taken from beneath every decorator: storage.UnwrapStore on the
direct routes, uow.UnwrapProvider on the proxied route. Over HTTP the
innermost store is the http client, whose annotation read is the server's
answer: against a server that does not advertise
policy.external_dependencies that answer is the column, and --raw is
served. A server that advertises it decorates the flag it answers, the v0
wire has no member asking it for the column, and the decorated flag cannot
be turned back into the column client-side, so --raw against such a server
is refused before the annotation read, naming the token, rather than
answered with the decorated flag. --raw without --flags is refused.

Tests: TestE2E_BlockedFlagsOverHTTP/raw_is_the_stored_column (served --raw
equals local --raw; the externally-blocked-only row reads true with --flags
and false with --raw on both; locally blocked rows agree; a stale-flagged
row reads true in both modes; --raw without --flags exits 1; a server
advertising the policy refuses --raw with no annotation request);
TestProxiedServerBlocked/flags_raw_is_the_stored_column; corpus
gc-blocked-flags-raw (pending until the reference server advertises
dependencies.blocking.isBlocked). Routing --raw through the decorated
annotator fails the e2e. CHANGELOG entry.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(changelog): bd blocked --flags --raw is for the local and proxied routes

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(blocked): give the e2e gate-kid its own priority; document --flags answers closed ids

TestE2E_BlockedOverHTTP seeded gate-parent and gate-kid at one priority, so
when both landed in the same second their order fell to the generated ids,
which differ between the local and the served workspace: the comparison
failed intermittently. gate-kid now has its own priority. Ten runs on the
embedded lane pass.

bd blocked --flags answers a closed id too, with its stored flag; a SQL read
of the column filtered on status <> 'closed' leaves it out. Said in the
command's doc and the CHANGELOG.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(blocked): apply label filters to external-only blocked rows

The external-dependency policy appends a row blocked only by an
unsatisfied external: ref after the label query has run, so --label,
--label-any and --exclude-label never judged it: bd blocked and
GET /v0/beads/blocked listed it whatever its labels, on every route.
Both decorators now filter those rows by the request's labels. The uow
decorator reads them through GetMany, chunked at MaxGetManyIDs, so their
labels are hydrated; GetIssuesByIDs left them empty, which also printed
them unlabelled on the proxied route.

Review fixes (adopt-pr apply-fixes, iteration 1):
- major #1: label filters on external-only rows in the store and uow
  decorators, with absence-asserting arms in the unit, served httpclient
  (server- and client-enforced policy), e2e and proxied tiers; the spec
  and the role doc say such rows are narrowed like every other row.
- minor #2: the policy.external_dependencies token prose names the
  blocked listing (spec, routes.go, regenerated types.gen.go).
- minor #6: SERVE_RUNBOOK and the allowUnlimited comment record the
  blocked listing's unbounded whole-set answer and why.
- minor #10: BlockedLister telemetry-outermost and typed-unsupported pins
  beside the ReadyCounter pair, plus the externaldeps wrap pin.
- nit #9: F-blocked names --flags --raw's refusal against a server that
  applies the policy (ledger.go and the regenerated ledger doc).
- nit #11: the remote BlockedLister no longer builds a local lister it
  then discards.

Review fixes (adopt-pr apply-fixes, iteration 2; the review approved at
940 with one minor and two nits, none gating):
- minor (the --flags refusal list was unpinned): the new
  TestBlockedFlagsRefusesEveryListingFilter pins blockedFlagFilterFlags
  to the flags blockedCmd registers, in both directions. Every
  registered flag except --flags and --raw must be refused, and every
  entry must name a registered flag. The test walks Flags() rather than
  LocalFlags(), so it does not merge the root's persistent flags into
  blockedCmd, and it skips inherited root flags and cobra's --help.
  Mutants:
  - four fail it: a new filter flag left off the list, a dropped
    entry, a misspelled entry, and --raw added to the list;
  - a simulated earlier in-process run (merged root flags plus --help)
    stays green, and fails once either skip is removed.
- nit (the remote branch's span): the remote BlockedLister comment now
  says why that lister is not wrapped again. The inner lister already
  carries telemetry's span, so wrapping would nest a second span of the
  same name. The comment also says what runs outside that span.
- nit (duplicated order comparator): left as is. finishBlockedIssues
  predates this PR, so merging the two sorts would refactor existing
  code with no review pass left after this one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-17 12:29:13 +00:00
Julian Knutsen
601be8a6bf
feat(list): bd list's remaining filters over http behind issues.list.filters (#7647)
* feat(list): bd list's remaining filters over http behind issues.list.filters

Over http every `bd list` filter beyond the created bounds and the title,
id and no-assignee scope refused (time bounds, priority range, text and external-ref
matches, spec, label pattern/regex, exclude-type, no-parent, no-labels,
empty-description, pinned/no-pinned, deferred, overdue, mol/wisp type).
The library entry already applied all of them: issueops.Reader.List via
workapi.BuildListFilter, shared by the CLI and the handler. This adds
only the wire.

- Server: GET /v0/beads/issues decodes 29 new optional parameters
  (listFilterParams, decode only; countIssues' names where it already
  spells the filter) into the existing ListRequest, behind one new
  behavior token, issues.list.filters. mol_type/wisp_type are closed
  enums refused by name (an empty value too). wire_revision stays 2;
  apigen and the wireshape golden are regenerated (additive only).
- Every RFC 3339 bound the listing and count decode is normalized to UTC,
  so an offset-bearing bound selects what its Z twin does on a zone-less
  timestamp column.
- workapi.BuildListFilter refuses PinnedFlag with NoPinnedFlag
  (ErrValidation, the CLI's own sentence) instead of letting PinnedFlag
  win; over http that is a 400 on no_pinned.
- Client: the encoder sends the members (encode.ListFiltersMembers, one
  list read by the encoder, the pre-dial gate and a spec-parity test).
  Against a server not advertising the token, httpReader.List refuses
  before dialing with ErrUnsupported naming issues.list.filters. The 29
  E-ListRequest ledger refusal rows retire; the ledger doc is updated;
  skew flag mappings added.
- Tests: handler decode/parity/enum/UTC/malformed/pinned-pair tests;
  encode member/gate/spec parity and bound encoding; per-member
  ride-behind-token and pre-dial skew refusal; two new Reader contract
  cases on every leg (dolt, embedded, uow, served): the eight time
  bounds at exact boundaries (strict, whole-second truncation, NULL never
  matches, +02:00 equals UTC) and one in/out row per remaining family;
  three served Reader cases unparked; TestE2E_ListFilterParityOverHTTP
  (served == embedded, masked server refuses naming the token, no dial);
  five corpus shapes marked pending until the reference server profile
  advertises the token.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(list): exclude_status on listIssues, the 30th issues.list.filters member

issueops.ListRequest gains ExcludeStatus: statuses to leave out of a
listing, appended to the default exclusions (never re-admitting what they
hide), comma-separated or repeated, and validated against the workspace's
statuses by the same rule CountRequest.ExcludeStatus uses (an unknown name
is ErrValidation). The rule moves into one shared function,
workapi.ParseExcludeStatus, called by both builders. With AllFlag it reads
the active set without enumerating every status, so an http-backed caller
can list O(active) rows instead of the whole history.

- Server: listIssues decodes a repeatable `exclude_status` under the same
  issues.list.filters token (no new token); an unknown name is a 400 on
  exclude_status. apigen and the wireshape golden are regenerated
  (additive).
- Client: the encoder sends it and it joins encode.ListFiltersMembers, so
  the pre-dial gate asks for the token.
- It is refused together with ReadyFlag, like the other members the ready
  query cannot carry.
- No `bd list --exclude-status` flag yet (it would need the CLI docs
  regenerated); the member is reachable through the library and the wire.

Tests: builder unit test (append, AllFlag, comma split, unknown name);
handler decode and 400; encoder, round-trip and gate; Reader contract
cases on every leg (active set under AllFlag, repeated and comma values,
adds to the defaults, narrows an explicit Status, unknown name refused).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(list): label_regex compile errors are 400s; review polish for list filters

- An invalid `label_regex` on GET /v0/beads/issues, /v0/beads/ready and
  /v0/beads/ready:count is now `400 invalid_argument` (`invalid_value` on
  `label_regex`) instead of `500 internal`. The database still compiles
  the pattern in its own dialect; a request that sent a regex and earned
  the engine's "regular expression is invalid" refusal is attributed to
  the parameter (failRegexReadErr). The same text on a request with no
  regex stays a server fault.
- mol_type and wisp_type refuse each other's vocabulary (tests pin
  `mol_type=heartbeat` and `wisp_type=swarm`), so the call-site
  allowlist, not just the spec enum, is covered.
- exclude_status with a workspace custom status: leg tests on the
  embedded store and over the served wire (accepted and applied; a name in
  no vocabulary refused). ReaderFixture cannot configure custom statuses,
  so these sit beside the count's twin rather than in the contract.
- The search e2e gains a positive arm: --desc-contains returns exactly the
  row whose description carries the token.
- Docs: `all` plus `exclude_status` keeps every status not named (pinned,
  hooked and custom done/frozen included unless named), not "the active
  set"; label_regex documents its 400. A corpus source line is made
  neutral; a test message says 30.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(list): compose issues.list.filters with the --title-contains fold

#7631, now on main, folds `bd list --title-contains` onto listIssues
`title` and refuses the TitleContains left over from a --title and
--title-contains pair that differs, because the listing then had one
title parameter. This branch publishes `title_contains` behind
issues.list.filters, so that leftover now rides it beside `title` and
the server applies both. Neither half is dropped, and against a server
without the token the pre-dial gate refuses, naming issues.list.filters.

- encode.FoldTitleContains still runs first; its doc now says where a
  differing pair goes.
- Tests: the fold table expects `title` plus `title_contains` for a
  differing pair, including the ſ/s pair strings.EqualFold would match;
  the differing-pair refusal names issues.list.filters; the
  list_filters_test sweep pairs its TitleContains probe with a differing
  TitleSearch, since a lone one asks only for issues.list.title; the e2e
  differing-pair refusal becomes a listFilterShapes row that answers
  exactly child-a over http and embedded, and refuses before dialing
  without the token.
- The gc-polecat-closed-after corpus shape's pending reason also names
  issues.list.title. The CHANGELOG --title-contains entry no longer
  lists the differing pair and --closed-after as refused.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* build(bazel): auto-sync BUILD files

Applied from the bazel-sync-patch artifact of https://github.com/gastownhall/beads/actions/runs/39013611268
(`make bazel-sync`: gazelle, tools/bazel/go_srcs.py, bazel mod tidy).
Only BUILD.bazel, MODULE.bazel and MODULE.bazel.lock are ever pushed; see
scripts/bazel-autofix-push.sh.

* fix(list): serve --priority-max in search parity, refuse the parent pair

The proxied tier's search parity table still expected a remote
`bd search --priority-max 1` to refuse, because listIssues could not
carry it. This branch serves priority_max behind issues.list.filters,
so the remote route answered and the dolt-race shard went red in
TestProxiedServerSearchRoutesAgree. The row now expects the answer on
every route. It was the last wireCan=false row; the branch stays, with
a comment, for the next flag listIssues cannot carry.

- ParentID with NoParent: a served listing now carries no_parent beside
  parent, and ANDing the two answered only the edge-less rows under the
  parent's dotted id. BuildListFilter refuses the pair with
  ErrValidation, in the sentence bd list and BuildCountFilter already
  use, so issueops.Reader.List refuses it on every backend and the
  listing answers 400 on no_parent. Pinned by a builder test, a reader
  conformance case (dolt, embedded, uow, served http) and two httpapi
  tests.
- POST /v0/beads/issues:claimNext shares the ready listing's
  label_regex but answered a pattern the database cannot compile with
  500. It now gives the listing's 400 invalid_value on label_regex,
  through the helper the listing's mapping is split into.
- The descendant walk's ledger rows said listIssues publishes no
  parameter for members it now serves behind issues.list.filters. They
  now say the walk does not send them yet, and that --flat or --json
  asks listIssues directly. The ExcludeStatus, Pinned and ExcludeTypes
  rows, the parentwalk.go header, the walk table's doc and two walk test
  comments are corrected the same way, and the ledger doc is
  regenerated. parentWalkRefusals notes that a member moved off the list
  must check issues.list.filters itself, because the walk dials past
  the listing's pre-dial gate.
- Spec: the listing's no_parent documents the pair's 400, the count's
  exclude_status no longer says it has no listing counterpart, and
  claimNext's label_regex documents its 400. types.gen.go is
  regenerated. Comment-only: countFilters, and readyScopeFields'
  ExcludeStatus entry.
- CHANGELOG and README say the default `bd list --parent` tree view
  does not carry the new filters yet; the label_regex entry names
  claimNext, and a Changed entry records the parent-pair refusal.

Second review pass: docs and comments only, no behavior change.

- The prose copies of the issues.list.filters member set now match
  ListFiltersMembers: CapIssuesListFilters and wire.CapListFilters name
  the status exclusion, and the spec's capabilities text names
  label-regex.
- Every time bound on the listing and the count goes through
  q.timestamp, which normalizes it to UTC, and sqlbuild compares it
  strictly at whole-second precision. listIssues' created_before and
  created_after, and countIssues' created, updated and closed bounds,
  now say so in the words the new listing bounds use. The count's
  updated bounds also say what they filter, and its closed bounds say
  that a NULL closed_at never matches. types.gen.go is regenerated.
- README and CHANGELOG: `--flat` or `--json` applies the filters to the
  parent's direct children, not its whole subtree.
- The 29 listRefusal rows this branch serves stay deleted rather than
  kept as retired rows, as #7550 deleted its TitleSearch, IDFilter and
  NoAssignee rows when it served them. The CHANGELOG now says the rows
  are removed, not retired.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
2026-10-17 11:21:18 +00:00
Julian Knutsen
925e26bc48
fix(http): serve bd list --title-contains by folding it onto title (#7631)
* fix(http): serve bd list --title-contains by folding it onto title

`bd list --title-contains X` refused against an HTTP workspace because
listIssues publishes no title_contains parameter. The store compiles
TitleContains and TitleSearch to the same predicate (LOWER(title) LIKE
'%v%'), so the client now folds TitleContains onto TitleSearch before the
pre-dial gate (encode.FoldTitleContains) when TitleSearch is empty or equal.
The request then rides the existing `title` parameter behind
issues.list.title, and against a server without that token it refuses before
dialing, naming the token. A --title and --title-contains that differ still
refuse: the listing has one title parameter and dropping either value would
widen the answer.

No wire change. The ledger row E-ListRequest.TitleContains narrows to the
differing pair.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(http): fold case-only title pairs and qualify the CHANGELOG

Review fixups for #7631.

- FoldTitleContains also folds a pair that differs only in letter case.
  The store binds each value as '%' + strings.ToLower(v) + '%', so a pair
  equal under strings.ToLower is one predicate. It does not use
  strings.EqualFold, which matches pairs such as "ſ" and "s" that lower to
  different patterns. Unit cases pin both directions, and the served test
  answers a case-only pair like the single flags.
- CHANGELOG: name what still refuses over HTTP: a pair that differs other
  than in case, either flag with --parent in the default tree view, and
  --closed-after, which gascity's polecat pre-flight lookup of recently
  closed beads (step 3d) uses.
- Corpus: add gc-polecat-title-contains-closed-after, the 3d lookup,
  pending until the reference server advertises issues.list.title and
  listIssues carries closed_after (E-ListRequest.ClosedAfter).
- Drop two assertions that could not fail (a request passed by value
  cannot be mutated), and bring the e2e shape header, the
  tokenedListFilters comment, the encoder table comment and the
  E-ListRequest.TitleContains ledger row (doc regenerated) up to date.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-17 06:00:33 +00:00
Julian Knutsen
a3e1b24e2b
fix(import): tolerate only typed dependency refusals, not walk-query errors (#7630)
* fix(import): tolerate only typed dependency refusals, not walk-query errors

With SkipDependencyValidationErrors set (the import mode), the dependency
pass recorded ANY error from checkHierarchy/checkCycle as a skipped edge
and continued. That included failed walk queries (an untranslated
statement, a statement timeout, a dropped connection), so an import could
commit without the edge and report the SQL error only as a skip reason;
on a backend whose transaction aborts on a failed statement, the import
failed later at an unrelated statement with the real cause hidden.

Tolerate only the typed refusals (ErrSelfDependency, ErrDependencyCycle,
*DependencyHierarchyConflictError) via isDependencyValidationRefusal and
return every other error, keeping the %w chain so serialization retry and
error classification apply.

Tests: sqlmock cases where the hierarchy and cycle walks fail under import
mode, on both the per-edge statements and the batch graph, must abort the
pass; typed refusals on both shapes are still skipped. Removing the
predicate fails both error tests.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(import): record the dependency-check fix in the CHANGELOG; tidy two test comments

Review fixups for #7630.

- CHANGELOG (Unreleased, Fixed): bd import now fails when a dependency
  check's own query fails, instead of reporting the SQL error as a skipped
  dependency and, on Dolt, committing without the edge. Refused edges (a
  self-dependency, a cycle, a hierarchy conflict) are still skipped. docs/
  is left alone because it tracks the pinned release (v1.2.2).
- create_test.go: two doc comments kept the line breaks left where the
  internal tracker IDs were removed. They are reflowed into sentences; no
  test logic changes.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-16 23:55:53 +00:00
Julian Knutsen
47d5db39ff
feat(http): bd heartbeat over bd serve (issues.heartbeat), and close the re-arm orphan-lease race (#7600)
* feat(roles): issueops.Heartbeater role on the dolt, embedded and uow legs

Adds issueops.Heartbeater.Heartbeat(HeartbeatRequest{Actor, IssueID}) ->
HeartbeatResult{IssueID, Holder, HeartbeatAt, LeaseExpiresAt, Rearmed},
the lease refresh behind `bd heartbeat`, as its own role with its own
accessor.

- One body: internal/storage/issueops.ExecuteHeartbeatInTx validates
  (*HeartbeatFieldError for a blank/over-long actor or empty id), refuses
  a wisp, then runs RenewLeaseInTx, which is HeartbeatIssueInTx's body
  reporting the renewal. HeartbeatIssueInTx is now a wrapper, so the legacy
  HeartbeatIssue legs answer the same text and sentinels.
- The refresh is fenced on the holder AND on the live claim: after the
  holder-fenced UPDATE, a read in the same transaction requires the issue
  to be in_progress under the holder (the `update leases set` prefix and
  the one IF() the Postgres dialect rewrites are unchanged; a Postgres port
  should expect more serialization retries from that read). A refusal deletes the caller's orphaned lease
  row, and the legs commit that refusal.
- A refresh writes only the lease row: no row_lock, version, updated_at,
  event, journal entry, Dolt commit or hook. A re-arm (a live hand-doled
  claim with no lease row the caller holds) is recorded as a mutation:
  row_lock reminted, updated_at stamped, a lease_rearmed event, a journal
  entry, a version, a Dolt commit (LeaseRearmVersionCommit) and on_update.
  The remint makes a re-arm collide with a racing release or close instead
  of leaving an orphaned lease on an open or closed issue (a pre-existing
  race on every Dolt route). Rearmed is true only when a row was created:
  a same-second refresh (0 changed rows on embedded Dolt) or a holder
  spelled differently refreshes the existing row in place and keeps its
  granted_at/granted_node. Reported times are rounded to whole seconds,
  the stored values; the binds stay unrounded so a same-second refresh is
  never a 0-row UPDATE on Dolt.
- Typed refusals: an absent id is ErrNotFound; a foreign holder is
  *LeaseConflictError{Holder} wrapping ErrAlreadyClaimed; a wisp
  (Ephemeral) or no live claim is *LeaseConflictError wrapping the new
  ErrLeaseNotHeld, which also matches ErrNotClaimable.
- Legs: DoltStore and EmbeddedDoltStore through the issue-operation
  funnel (DOLT_ADD of the named tables, then DOLT_COMMIT -m, after the SQL
  commit); uow.NewHeartbeater through the new uow.RunTxScoped: a refresh or
  a refusal commits with the plain SQL COMMIT (no working-set scan, no Dolt
  commit, nothing another writer left uncommitted is swept), and a re-arm
  commits through the new ScopedCommitter (DOLT_ADD of exactly its tables,
  DOLT_COMMIT -m, never -Am). The legacy HeartbeatIssue doors commit a
  re-arm the same way and keep a refusal's orphan delete, and the
  proxied-server `bd heartbeat` route runs the role's body with the same
  scoped commit.
- Census: storage.Storage gains Heartbeater(); HookFiringStore wraps it
  (fires on_update for a re-arm only, RoleFiresHooks knows it), telemetry
  wraps it, httpapi.timedProvider, uow.notifyingProvider (records a re-arm)
  and externaldeps.uowProvider build it; the journal and version guards
  list the heartbeat family as entry points; the httpclient unsupported
  shell is regenerated and Heartbeater sits on its unsupported allowlist
  until the wire lands.

Conformance: RunHeartbeater* (18 cases) on dolt, embedded dolt and uow:
refresh post-state with the stored times pinned to the result; no
row_lock/current_revision/updated_at/events/journal/issue_versions/
version-control/hook change; a guarded write with the pre-heartbeat
revision still lands; foreign holder refused against an aged lease;
refused after close/release/reclaim/status move; wisp refused; hand-doled
claim re-armed as a mutation; same-second heartbeat is not a re-arm;
orphaned leases on released, closed and reassigned issues refused and
deleted; a refresh and a refusal against a dirty working set make no
commit and leave the other writer's change uncommitted, and a re-arm
commits only its own tables; absent id; request rules; no request mutation. The http leg is
waived until the wire lands.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(httpapi): POST /v0/beads/issues/{id}:heartbeat behind issues.heartbeat

Publishes issueops.Heartbeater on bd serve as heartbeatIssue, the served
lease refresh behind `bd heartbeat`.

- Spec: the operation, HeartbeatIssueRequest {actor} and
  HeartbeatIssueResponse {id, holder, heartbeat_at, lease_expires_at,
  rearmed}; no `revision` (a refresh writes nothing to the row) and no
  expected_version (the fence is the holder and the live claim).
  issues.heartbeat joins the capability vocabulary and lease_not_held
  (409) the Problem.code list; `reason: ephemeral` qualifies it for a
  wisp. apigen and the wire-shape golden are regenerated; the golden diff
  is additive only, so wire_revision stays 2.
- Server: route row on the custom-method dispatcher, handleHeartbeat in
  heartbeat.go (claim's actor rules via bodyActor, unknown member and
  query refusals), Config.Heartbeater as a required role (sourceRoles,
  roleSourceNames, the per-request uow.HeartbeaterSource accessor).
  failHeartbeat maps *LeaseConflictError typed fields: a foreign holder
  is the claim's conflict shape (already_claimed + assignee +
  issue_status); no live claim, a wisp included, is the new
  lease_not_held (+ issue_status, + reason: ephemeral for a wisp), matched
  before the generic arms so ErrLeaseNotHeld's ErrNotClaimable never
  surfaces as not_claimable. Its detail names the one state the refusing
  transaction found (closed, open and not claimed, an ephemeral wisp, ...).
- bd serve wires the role from its peeled store (serveRoleSource gains
  Heartbeater); the serve stub stores, the corpus and e2e harnesses, the
  served harness and the encode oracle bind it.
- Wire client vocabulary: OpHeartbeatIssue/issues.heartbeat and
  MethodHeartbeat are CLASSIFIED (the capability set-equality gate),
  ReasonEphemeral is pinned to the server's, and lease_not_held rebuilds
  *LeaseConflictError{Status, Ephemeral} wrapping ErrLeaseNotHeld in the
  native words (the code-table set-equality gate). No accessor yet.

Tests: heartbeat_test.go (request mapping and projection, edge refusals,
every role refusal with its code, members and state-specific detail, the
code row, the capability). The served-http Heartbeater leg
(served_heartbeater_test.go, 18 cases) runs the contract against a real
bd serve through a test-local dial over the store's own dispatch, and the
http-leg waivers are deleted.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(http): bd heartbeat over http through httpclient.Heartbeater

The client leg of the served lease refresh, and `bd heartbeat` on the
role for every route.

- httpclient.Store.Heartbeater (heartbeater.go) dials POST
  /v0/beads/issues/{id}:heartbeat through Store.dispatch, so
  issues.heartbeat is preflighted: an older server is refused BEFORE the
  dial with *wire.CapabilityError naming the token (unwraps to
  *storage.ErrUnsupported{Capability: "issues.heartbeat"}); it never falls
  back to a re-claim or a metadata stamp.
- The role's request rules are restated before the dial
  (validateHeartbeatRequest, pinned to
  storeops.ValidateHeartbeatRequest): blank/over-long actor, empty id, each
  a *HeartbeatFieldError.
- Problem mapping: already_claimed (which the shared table rebuilds as the
  claim's ClaimConflictError) becomes *issueops.LeaseConflictError naming
  the holder and wrapping ErrAlreadyClaimed, in the native words ("issue
  already claimed by <holder>"); lease_not_held already arrives as
  LeaseConflictError (Ephemeral from `reason`); not_found is ErrNotFound;
  invalid_argument on `actor` is the role's field error.
- Idempotent: true, the one write this client marks replayable (a repeat
  only moves the expiry later; a replay after a reclaim fails the fence).
- Legacy Store.HeartbeatIssue rides the role. Both leave the generated
  shell and completeness_test's unsupported allowlist.
- `bd heartbeat` takes Heartbeater() from the resolved store on the direct
  routes (dolt, embedded, http) and from the provider's HeartbeaterSource
  on the proxied route (exact id resolution in a read unit of work, then
  the role); runHeartbeatProxiedServer and heartbeat_proxied_server.go are
  deleted. Success output and exit codes are unchanged; refusal text
  changes per route (see CHANGELOG).

Tests: heartbeater_test.go (stub server: preflight refusal naming the
token with no POST, request/answer projection, each refusal's typed
mapping including the wisp's reason, the 503 replay, pre-dial validation,
validator parity, the legacy door). served_heartbeater_test.go runs the
18-case contract through env.subject.Heartbeater(), with
HookedHeartbeater wired off a HookFiringStore. TestE2E_HeartbeatOverHTTP:
real bd against an in-process serve: holder ok with revision and
updated_at unchanged and a guarded write with the pre-heartbeat revision
landing; a non-holder refused against an aged lease that stays untouched;
closed, wisp and unclaimed refused, each naming its own state; an older
server's handshake refused naming issues.heartbeat with no write. Corpus
shapes gc-heartbeat, -non-holder, -closed, -wisp pending until
reference_server.json advertises the token. CHANGELOG entry for the whole
track, including the orphaned-lease race fix.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* build(bazel): auto-sync BUILD files

Applied from the bazel-sync-patch artifact of https://github.com/gastownhall/beads/actions/runs/38875641041
(`make bazel-sync`: gazelle, tools/bazel/go_srcs.py, bazel mod tidy).
Only BUILD.bazel, MODULE.bazel and MODULE.bazel.lock are ever pushed; see
scripts/bazel-autofix-push.sh.

* fix(heartbeat): roll back a no-issue refusal; restate refresh vs re-arm

Review fixes for #7600 (adopt-pr apply-fixes, iterations 1-2).

Legacy HeartbeatIssue doors commit a refusal only when it read the issue.
The dolt and embedded HeartbeatIssue callbacks committed every
LeaseRenewalRefusal, including Found=false (the issue read failed). That
refusal deletes nothing, but it can follow a fast-path UPDATE that
already extended an orphan lease row's expiry, so committing it kept the
extension. issueops.LeaseRenewalRefusalCommits now gates the commit on
Found, the rule the Heartbeater role's HeartbeatRefusalCommits applies.
New tests on both legs pin both arms: a no-issue orphan's expiry is
unchanged, and a released issue's orphan is deleted.

Refresh vs re-arm contract docs. Comments, Long help, the OpenAPI spec
and the generated apigen still said every heartbeat writes only the
lease row, makes no Dolt commit and fires no hook. A re-arm is recorded
as a mutation: it remints the revision, records an event and a version
entry, commits once (LeaseRearmVersionCommit) and fires on_update
through the wrapping hook decorator. Restated in domain, uow, issueops,
storage.go (why RoleFiresHooks answers true, so bd serve takes the
heartbeater off a peeled store), cmd/bd heartbeat and serve, httpapi,
httpclient, the spec (apigen regenerated), and the conformance and
contract-test docs.

Minor review findings:
- The 409 already_claimed and lease_not_held details now end "nothing
  on the issue was written": the refusal may have deleted and committed
  the caller's orphan lease row.
- docs/multi-agent/federation.md and the CHANGELOG entry describe the
  spelling-mismatch heartbeat: the row is refreshed in place and keeps
  granted_at and granted_node; only a re-arm stamps this node. The
  embedded spelling test pins both columns, and the lease.go comments
  say "refreshed in place" instead of "re-arms".
- The refusal's orphan delete judges the holder under actorMatches, as
  the in-place refresh does (iteration 2). It matched holder = ? on the
  actor as sent, so a dead claim's orphan held under another spelling
  of the caller (gastown.mayor, heartbeat as gastown__mayor) survived
  the refusal until TTL. The refused closure now reads the row's holder
  and deletes that stored holder only when actorMatches(holder, actor).
  Both TestHeartbeatIssueCommitsARefusalOnlyWhenItReadTheIssue twins
  gain a cross-spelling arm (row deleted) and a control: another
  identity's row (gastown.dog-3) survives.

Not changed: `bd heartbeat --json` still omits `rearmed`, since adding
it changes a CLI JSON contract (follow-up). The federation.md paragraph
post-dates the v1.2.2 docs pin, so documenting this main-only behaviour
there is a maintainer call under
engdocs/decisions/2026-07-17-docs-release-pin.md section 1.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(uow): run the scoped re-arm's Dolt commit after the SQL commit

The proxied-server heartbeat commits a re-arm with doltServerTx.CommitScoped,
which ran DOLT_ADD per table and DOLT_COMMIT('-m') INSIDE the open
transaction, before the SQL COMMIT. DOLT_ADD stages whole tables from the
session's transaction root, so that Dolt commit was built from the re-arm's
start snapshot: an issues row another session changed while the re-arm was
open (a write still only in the working set, as in batch mode or before that
writer's own Dolt commit) was staged at its old value, and HEAD disagreed
with the rows every reader sees. This is the LatentLabsSpace/NEXUS#92
ordering the DoltStore funnel (runIssueOperationTxWithMessage) already
avoids, and the review's major finding on this PR.

CommitScoped now commits the SQL transaction first, then stages exactly the
named tables (plus the versioned-history tables) and makes the Dolt commit on
the same pinned connection, the uow twin of doltAddAndCommitPostTx:

- a serialization failure of the SQL COMMIT is still returned, so
  RunTxScoped replays the unit of work;
- the Dolt commit runs detached on its own budget and replays only conflicts
  that leave nothing behind: serialization failures and Dolt's 1105
  "@autocommit transaction rolled back" (new isDoltAutocommitRollbackError,
  mirroring the DoltStore classifier of the same name);
- nothing staged, or "nothing to commit", is not an error: a concurrent
  writer's Dolt commit already took the row. A "nothing to commit" past the
  staged check is logged, as the funnel logs it, so the missing audit line
  is explicable;
- any other failure is logged, never returned. The data is committed, and
  returning it would make RunTxScoped replay a committed re-arm, which would
  then report a refresh. The session is poisoned rather than pooled, since
  its state is unknown. uow records no metrics, so that log line is the
  drop's only signal; bd.db.post_tx_commit_dropped counts only the funnel's.

Statements go through schema.DrainCall, retiring the local drainCall fork
(review nit).

Tests: five sqlmock tests pin the order (SQL COMMIT before any DOLT_ADD),
the swallow-and-poison path, the 1105 retry, the returned serialization
failure and the empty-commit tolerance; each fails against the previous
CommitScoped. TestIsDoltAutocommitRollbackError carries the DoltStore
classifier's table over to the uow copy, plus the prefix boundary; dropping
the prefix branch, the prefix's period, the 1105 check or the type check
each fails a case. TestHeartbeaterRearmKeepsAConcurrentWrite is the
real-Dolt regression: a change to another issues row made while the
re-arm's transaction is open must be at HEAD after the re-arm's Dolt
commit. Its working-set-only writer fails on the old ordering (dolt
2.1.8); the writer that Dolt-commits its own change passes on both and is
kept as coverage.

Review follow-ups folded in (comments, a file move and a test; the
absorption log line above is the only behaviour change among them):

- the freshRowLock invariant and its row_lock_guard_test.go mirror name the
  re-arm as the one deliberate remint without a lifecycle-field change,
  load-bearing for the release/close collision;
- httpclient: "a write is never retried" (wire/client.go, stall_retry.go)
  and "only a read's answer is yes" (dispatch.go) now allow for the
  Idempotent heartbeat, and httpHeartbeater and issueops.HeartbeatResult
  document that a replayed heartbeat can report Rearmed false after a
  committed re-arm, with the guarded write's revision check as the backstop;
- the proxied route moves back to cmd/bd/heartbeat_proxied_server.go, the
  file family cmd/bd/AGENTS.md prescribes;
- the ScopedCommitter and RunTxScoped docs no longer say every other
  writer's uncommitted change stays out of the commit: a staged table is
  committed whole, so such a change to a named table rides in the re-arm's
  commit, as it does in the funnel, and only other tables keep theirs;
- commitStagedTables says its drop is log-only, the absorption is logged,
  and the uow classifier gets its table test (see Tests).

Not folded in: the base-owned IssueUseCase.Heartbeat /
IssueSQLRepository.HeartbeatIssue door, dead in production since the
proxied route moved to the Heartbeater role. Its one caller is the
notifying hook-parity test (uow/notifying_test.go), and moving that case
onto the Heartbeater means re-deriving its RenewLease hook semantics, so it
is left for a follow-up.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-16 22:38:23 +00:00
gascity-eddie[bot]
91a72ef7b5
fix(mol): route bond operand resolution like show/update/close (closes #4714) (follow-up to #4720) (#7622)
* fix(mol): route bond operand resolution like show/update/close (closes #4714)

bd mol bond resolved operands with utils.ResolvePartialID against the local
store only, so a bead that lived in a routed store failed with
"'<id>' not found (not an issue ID or formula name)" — even though bd show
resolved the same ID. This broke `gt sling <bead> <rig>`, which cooks the
default formula and bonds it to a rig-routed bead.

- Resolve operands through resolveAndGetIssueForMutation, the same routing
  fallback bd show/update/close use (local, then prefix-routed rig via
  routes.jsonl, then contributor auto-routing). Write-intent so a prefix-routed
  target opens writable and the bond commits where the bead lives.
- Run the bond mutation against the store that owns the resolved issue instead
  of the global local store; reject a bond whose two operands resolve to
  different stores rather than writing to the wrong database.
- Route the --dry-run preview read-only so it previews routed operands too.
- Add TestResolveMolBondOperandRouting covering local and routed resolution.

Same routing-parity fix as #3608/#3609 for bd close.

* fix(mol): fail fast on read-only routed stores, share same-rig bond handles

Address the two structural issues from the #4720 review beyond the rebase
itself (the rebase already preserves main's unconditional formula fallback
from #3874 instead of reintroducing the looksLikeFormulaName gate):

- Routed operands resolve but couldn't be bonded: contributor auto-routing
  deliberately opens stores read-only (they hydrate foreign projects that
  must never be mutated), yet the bond selected that store and dispatched
  mutations that only failed at the store layer. Operand resolution now
  mirrors resolveCloseTargets explicitly (local, then prefix-routed
  write-intent, then the shared auto-routed read-only handle) and records
  writability per operand; pickBondStore rejects a bond pinned to a
  read-only store with a clear message before any mutation is attempted.
  TestMolBondPrefixRoutedWriteThrough now verifies the actual dependency
  write commits in the routed database, not just resolution, and
  TestMolBondAutoRoutedReadOnlyFailsFast covers the fail-fast path.

- Same routed database compared as two different stores: pickBondStore
  compared Go handle identity, but each route-open constructs a fresh store
  in server mode, so two operands in the same routed rig were wrongly
  rejected as cross-store. Both operands of a bond now share one routed
  handle (an already-opened routed store is tried before opening another),
  and sameBondStore falls back to StoreLocator identity (CLIDir pins the
  concrete database) for separately-opened handles onto the same database.
  TestSameBondStoreSeparateHandles covers the two-handle case directly.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(mol): validate routed bond homes before mutation

Discover bond operands read-only, enforce dry-run parity and cross-store policy before writable reopen, unwrap decorated store identities, preserve actionable route failures, and verify the accepted home after reopen.

Co-authored-by: Johan Snyman <johan.snyman@orange.com>

Agent-Signature: codex-gpt-5.6-sol-high on behalf of matt wilkie

* fix(mol): reconcile routed bond discovery with #5253 var validation

The rebase onto current main carried two silent regressions the Aug 16
review named: the discovery-based resolution (which replaced
resolveOrDescribe on the direct-store path) never validated --var values,
so `bd mol bond --dry-run <formula> <x> --var bad=...` previewed success
for a bond execution refuses — exactly the false preview #5253 fixed —
and materializeMolBondOperand wrapped every cook error as
"'%s' not found as issue or formula", re-introducing the ErrVarValidation
double-wrap #5253 removed from resolveOrCookToSubgraph.

discoverMolBondOperand now takes the provided vars and runs
formula.ValidateProvidedVars on its formula fallback, so --dry-run and
execution reject the bond identically, before any writable reopen.
materializeMolBondOperand passes formula.ErrVarValidation through
unwrapped, matching resolveOrCookToSubgraph and pour/wisp.

Agent-Signature: claude-fable-5 on behalf of R0SEWT

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(mol): review polish — fs.ErrNotExist, hook-isolated test repos, auto-route coverage

Three non-blocking notes from the Aug 16 review:

- routed.go: errors.Is(err, fs.ErrNotExist) instead of the legacy
  os.IsNotExist, so the not-found check keeps working if loadPrefixRoutes
  ever starts wrapping its open error.
- mol_bond_routing_test.go: pin core.hooksPath to .git/hooks after each
  git init (the initGitRepoAt convention), so a developer or CI box with a
  global hooksPath override never runs foreign hooks inside test repos.
- The widened openAutoRoutedStore behavior (maintainer/default auto-routes
  open writable for any validated write-intent resolver, not just mol bond)
  now has both the CHANGELOG sentence and a write-through test:
  TestMaintainerAutoRouteMutationWritesThrough drives
  resolveAndGetIssueForMutation — the bd update resolver — through a
  maintainer auto-route, commits a real update there, and re-reads it via a
  fresh routed read.

Agent-Signature: claude-fable-5 on behalf of R0SEWT

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(mol): carry main's #5761 formula validation into routed bond discovery

Rebase adaptation for #4720 onto main b303d494df. The PR moved the direct
(non-proxied) `bd mol bond` path off resolveOrDescribe and
resolveOrCookToSubgraph onto discoverMolBondOperand and
materializeMolBondOperand. Main's #5761 meanwhile fixed the old pair so a
--dry-run fails on exactly what the real bond fails on; the new pair did not
have that fix:

- discoverMolBondOperand loaded the formula with parser.LoadByName, which
  never runs Formula.Validate, so a --dry-run previewed a successful bond for
  a formula the cook rejects (e.g. a waits_for gate with no spawner). It also
  validated --var values against the unmerged formula, so a var whose enum is
  declared only in a parent went unchecked. Discovery now resolves the
  formula with parser.Resolve first, the way resolveAndCookFormulaWithVars
  does, and wraps both failures the way the cook does, so --dry-run and
  execution report the same message.
- materializeMolBondOperand passed only formula.ErrVarValidation through
  unwrapped. A formula that fails validation was reported as "not found as
  issue or formula". formula.ErrValidation now passes through too, as in
  main's resolveOrCookToSubgraph.

TestDiscoverMolBondOperandRejectsWhatMaterializeRejects and
TestDiscoverMolBondOperandValidatesVarsAgainstTheMergedFormula assert the two
phases against each other, like main's TestBondDryRun* pair does for the
proxied route.

Also:
- The PR's routing tests now pass the `exact` argument that main's #5393
  added to resolveViaPrefixRoutingWithAccess. They pass false, which keeps
  the partial-ID resolution the tests were written against.
- The PR's three new cgo test files are added to the bd_test srcs. This is
  `make bazel-sync` output on the merged tree, and it touches nothing else.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Rody Vilchez <102562850+R0SEWT@users.noreply.github.com>

* test(mol): rebuild the RepoContext cache in the PR's role-routing tests

Rebase adaptation for #4720 onto main b303d494df. Main's #5243 moved
auto-routing role detection onto beads.GetRepoContext(), which caches the
repo context once per process; before it, role detection ran
routing.DetectUserRole(".") on every call. The PR's auto-routing tests chdir
into a fresh repo and set beads.role there, so after the rebase they read
the context an earlier test built for its own, now-deleted repo. There
beads.role is unset, and the remote-URL fallback finds no origin and answers
maintainer:

- TestMolBondAutoRoutedReadOnlyFailsFast fails whenever a role-routing test
  ran before it in the same binary (reproducible alone with -count=2):
  discovery looks for the contributor operand in the local store and reports
  it not found.
- TestDiscoverMolBondOperandRouting/maintainer_routed_issue and
  TestMaintainerAutoRouteMutationWritesThrough still pass, but only because
  that fallback is also maintainer; they never read the beads.role they
  configure.

Each test now calls resetRepoContextCachesForTest after its chdir, as #5243
did for main's own role-routing tests (TestResolveCloseTargets, which
TestDiscoverMolBondOperandRouting mirrors, and the routing_read tests).

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Rody Vilchez <102562850+R0SEWT@users.noreply.github.com>

* fix(routing): open maintainer auto-routes writable for exact-match writes

#5393 landed after this PR was written. It moved `bd comment` and
`bd comments add` off resolveAndGetIssueForMutation onto a new
resolveAndGetIssueForMutationExact. This PR widens the former's auto-route
tier for write intent: a maintainer/default target opens writable, while a
contributor target stays read-only and MutationForbidden. On the PR's
original base that widening covered comment writes. After the rebase the
exact resolver still opened every auto-route read-only, so comment writes
lost it and the CHANGELOG's "any validated write-intent resolver" no longer
held.

The exact resolver is meant to keep the mutation resolver's routing
fallbacks and only tighten matching (#5393: it "keeps the existing
cross-rig prefix-routing and contributor auto-routing fallbacks intact").
Its auto tier now takes the same write-intent path. Matching stays exact,
and the abbreviation-refusal precedence is unchanged.

TestMaintainerAutoRouteExactMutationWritesThrough pins the change: the
routed store must not be read-only, and a comment written through it must
commit in the planning database. Without this change the read-only check
fails. The write itself still succeeds, because a server-mode read-only
open does not refuse it.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Rody Vilchez <102562850+R0SEWT@users.noreply.github.com>

* fix(routing): harden routed writes and surface routing failures

Review fixes for #4720.

bd close resolved its batch's shared auto-route through
openRoutedReadStore, so the route opened read-only even though every
other write-intent resolver opened it writable. The shared handle now
opens through openAutoRoutedStore with write intent, so the close commits
where the issue lives.

A contributor auto-route now opens writable for write-intent callers, as
maintainer and default routes do. The refusal the previous revision added
(errContributorRouteMutationForbidden, contributorRouteRefusal,
RoutedResult.MutationForbidden) is gone. A contributor's route target is
their own planning repository, the one bd create writes to:
docs/multi-agent/multi-repo-migration.md and
examples/contributor-workflow/README.md have the contributor claim and
close planning issues from the fork, and GH#2345, beads-0km and GH#3608
made bd show, update and close reach those issues. The read-only rule
came from #4389's review fixes (e8ae7a291a), which kept "contributor
auto-routing" read-only to preserve the bd-6dnrw.32/GH#3231 guarantee
that "routed reads never mutate a foreign project's history". That
guarantee is about reads, and the planning repository is not a foreign
project. Base therefore split by backend: a server-mode bd update, close
or comment of a planning issue committed, because a read-only server
store accepts the write, while the embedded store refused it with "store
is read-only". Refusing in every resolver would have broken the
documented workflow on both backends. bd mol bond was the only command
that checked MutationForbidden, so its check goes too, and a bond of
planning issues commits in the planning repository like any other routed
write. Reads through every route stay read-only. This reverses the
previous revision's stance and is flagged here as a maintainer call.

bd unclaim now resolves its issue through resolveAndGetIssueForMutation,
as bd update --claim does, so a routed issue opens writable; with the
read resolver an embedded routed target refused the release with "store
is read-only". Like bd close, it commits a routed change itself
(commitPendingIfEmbedded), since the post-run auto-commit covers only the
local store. bd human respond and dismiss already opened their target
writable but never committed it, so on an embedded workspace the close
sat uncommitted in the routed repository's working set; they now commit
it the same way. resolveAndGetIssueWithRoutingAccess's
writablePrefixRoute parameter is renamed writable, since it now governs
auto routes too.

A routed commit that fails no longer looks like success, and a retry
commits it. bd unclaim printed "Unclaimed" (or listed the issue in its
--json output) even when the release's routed commit failed; it now
reports the commit error and leaves the issue out. The failed commit
leaves the release, or the close human respond and dismiss make, in the
routed repository's working set. There a retry finds the issue
unassigned or already closed and refuses before it reaches its commit,
and nothing else commits a routed target. On that refusal unclaim,
respond and dismiss now commit the pending write before reporting it,
through commitRoutedWrite, which their success paths share.

A routed write never creates the target's database. The routed open runs
before the lookup it serves, so an embedded open of a repository that had
not been bootstrapped yet created an empty database that outlived an id
that missed; bd bootstrap then took it for the real one and skipped the
clone from sync.remote. openRoutedTargetStore's writable branch now opens
through newExistingDoltStoreFromConfig, which refuses an embedded
workspace without a database (embeddeddolt.HasRepository) before the
sanitize migration can rewrite anything there. A server-mode open already
never creates one.

resolveAndGetIssueForMutation already returned a routing failure (an
unreadable routes.jsonl, a routed store that fails to open) as itself. The
exact resolver and bd close still fell through to "no issue found
matching", which misstates why the lookup missed. All of them now share
isRoutingFailure. The exact resolver checks its abbreviation refusal first,
so the auto tier still gets its chance at an exact match. A prefix route
whose target has no usable metadata.json now names the route and the file
and says why the file could not be used.

An auto-routed store now opens the target's own dolt_database, as a prefix
route already did. Otherwise BEADS_DOLT_SERVER_DATABASE, which names the
launching workspace's database, would point a server-mode open at the wrong
database on a shared server. openRoutedTargetStore holds that override for
both route kinds. A write refuses a target whose metadata.json cannot be
read or parsed or names no dolt_database (autoRouteTargetDatabase) instead
of falling back to that variable. A target with no metadata.json at all
opens, read or written, the only embedded database it holds
(embeddeddolt.Databases), or the default database
(configfile.DefaultDoltDatabase) when it holds none or several.
Contributor setup leaves the planning repository without one until the
first routed bd create writes it, and names its database the way any
workspace without metadata.json does, BEADS_DOLT_SERVER_DATABASE first.
The open would otherwise resolve that name again from the launcher's
variable, so when the two differed a read that missed failed with
"database not found" and a write that missed created a database beside
the one setup created. A server-mode target holds no embedded database,
so with the variable set such a route now opens the default database
rather than the launcher's, which is what it opens without the variable.
A read of a target whose metadata.json is unreadable or names no
database keeps the lenient lookup it had.

bd mol bond's routed home now gets the external-dependency policy and
nothing else, as wireExternalDependencyPolicy asks of routed stores. The
launching workspace's hooks and telemetry no longer wrap writes into
another rig. runMolBond's tail is split into planMolBond, openMolBondHome,
materializeMolBondOperands and bondMolBondOperands, which takes it from
cyclomatic 26 down to 10 or less per function. getIssueWithRouting and the
resolveViaAutoRouting wrapper had no callers left and are removed.

Tests:
- TestMaintainerAutoRouteMutationWritesThrough now checks that the routed
  store is writable. Without that check it passes even against a read-only
  open, because a server-mode read-only store accepts the write.
- TestAutoRouteCloseTargetWriteIntent covers bd close's write intent for a
  maintainer route.
- TestContributorAutoRouteWritesThrough replaces
  TestContributorAutoRouteRefusesWriteIntent: an update, a comment and a
  close through a contributor route each open the target writable and
  commit, a read through the same route stays read-only, and a missing id
  stays a plain not-found.
- TestEmbeddedContributorRoutedWritesCommitInPlanning runs bd update
  --claim, unclaim, comment, close, human respond and human dismiss
  through the bd binary in a contributor checkout. Each advances the
  planning database's HEAD, and the issues end up closed there and absent
  from the fork's database. The embedded shard manifest gains the test.
- TestMolBondContributorAutoRoutedWriteThrough replaces
  TestMolBondAutoRoutedReadOnlyFailsFast: bd mol bond's pipeline bonds two
  planning issues through a contributor route, and a fresh routed read
  sees the dependency; setupAutoRoute now seeds extra planning issues for
  it. TestDiscoverMolBondOperandRouting's contributor case now expects the
  writable reopen.
- TestResolveCloseTargets keeps base's contributor cases (beads-0km); the
  previous revision had moved them to routing.maintainer.
- TestRoutedWriteDoesNotCreateTargetDatabase (in-process) and
  TestEmbeddedRoutedWriteMissDoesNotCreateTargetDatabase (the bd binary)
  check that a missed write through a prefix route and through a
  maintainer auto-route leaves an embedded target that has not been
  bootstrapped without an embeddeddolt directory. The embedded shard
  manifest gains the test.
- TestRoutedMutationResolversReportRoutingFailures runs each write-intent
  resolver against an auto route whose store fails to open, an auto route
  whose target's metadata.json names no dolt_database while
  BEADS_DOLT_SERVER_DATABASE names the launcher's, and a prefix route
  whose target cannot load.
- TestAutoRouteTargetDatabase covers a write and a read of a target whose
  metadata.json names a database, is missing, names none, or cannot be
  parsed; a missing one yields the target's only embedded database, or
  the default database when it holds none or several, either way.
  TestDatabases covers the listing, which applies HasRepository's .dolt
  marker probe to each database.
- TestAutoRouteMissWithoutMetadataIgnoresEnv runs a missed read and a
  missed write through an auto-route to a target that has one database
  but no metadata.json: the default database while
  BEADS_DOLT_SERVER_DATABASE names the launcher's, and a database named
  after the launcher's with the variable set and unset at use. Each miss
  must report not found and leave the target holding only its database.
- TestRoutedWriteRetryCommitsFailedCommit fails the routed commit of bd
  unclaim (text and --json), human respond and human dismiss with an
  invalid --dolt-auto-commit. The run must fail with nothing on stdout and
  leave the routed repository's HEAD as it was; a retry must report its
  refusal and commit the write there.
- TestAutoRouteOpensTargetDatabaseDespiteEnv checks the target-database
  change.
- The auto-routing tests clear BEADS_ROUTING_MODE and BD_ROUTING_MODE
  through neutralizeRoutingModeEnv. Viper ranks them above the
  routing.mode the tests write, so a host that exports either var never
  routes. bdEnv, which builds the bd binary's environment in the
  embedded tests, now drops BD_ROUTING_MODE as it drops BEADS_*; agent
  hosts export it, and TestEmbeddedContributorRoutedWritesCommitInPlanning
  failed on such a host.
  TestEmbeddedRoutedWriteMissDoesNotCreateTargetDatabase no longer needs
  its own envWithout.
- Reverting each fix turns its test red: the contributor read-only clamp
  (the contributor tests above), unclaim's resolver or commit and the
  respond and dismiss commits (the matching step of the embedded test),
  and each autoRouteTargetDatabase rule. So does dropping unclaim's,
  respond's or dismiss's retry commit, or unclaim's skip of a release
  whose commit failed (TestRoutedWriteRetryCommitsFailedCommit); either
  earlier no-metadata.json rule, the variable, which leaves the stray
  database the review reported, or a fixed default database, which fails
  the read of a target setup named after the variable
  (TestAutoRouteMissWithoutMetadataIgnoresEnv and
  TestAutoRouteTargetDatabase), and taking the first of several databases
  (TestAutoRouteTargetDatabase); and bdEnv's BD_ROUTING_MODE drop, on a
  host that exports BD_ROUTING_MODE=off (both embedded tests).

The CHANGELOG entry now records that every auto-route opens writable for
the commands that change an existing issue, the unclaim and human
respond/dismiss commits, that a routed write never creates the target's
database, and the target-database and routing-failure changes.
docs/multi-agent/routing.md says that write commands follow every route,
contributor, maintainer and default alike, never create the routed
repository's database, and refuse a target whose metadata.json is
unreadable or names no dolt_database. Both say that an auto-route opens
a target with no metadata.json at the one embedded database it holds, or
its default database when it holds none or several, while a prefix route
always needs the target's dolt_database. That page documents
behavior that is not in a release yet, which the docs release-pin
decision (engdocs/decisions/2026-07-17-docs-release-pin.md §1) would
leave undocumented. The review asked for docs that match the behavior,
as feature PRs #6508, #5855, #5997 and #5092 also did, so it is flagged
here as a maintainer call.

Open, and not changed here. bd close has the same retry gap on base: a
re-close of an issue that a failed routed commit left closed counts as
already closed and never commits that store (mutatedStores).
commitPendingIfEmbedded gates on the launcher's mode (isEmbeddedMode),
not the routed store's, so a server-mode launcher that opens an embedded
routed repository would leave the write uncommitted there. That gate is
shared by every routed write command and is not exercised here.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Rody Vilchez <102562850+R0SEWT@users.noreply.github.com>

* ci(shards): list TestDatabases in the 40-shard embedded storage manifest

The routed-write hardening added TestDatabases to
internal/storage/embeddeddolt/detection_test.go without a slot in the
Bazel-only 40-shard block of .github/scripts/embedded-storage-test-shards.txt,
so gen_embedded_storage_shard_manifest.py 40 --weights=duration --check
reported it missing and TestStorageEmbeddedShardManifestGeneratorNotStale
(go test ./scripts) failed.

Regenerated with the repository command
(gen_embedded_storage_shard_manifest.py 40 --weights=duration --write);
the incremental pack adds exactly one line, on shard 25.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Rody Vilchez <102562850+R0SEWT@users.noreply.github.com>

---------

Co-authored-by: Rody Vilchez <102562850+R0SEWT@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: matt wilkie <maphew@gmail.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <338826946+gascity-eddie[bot]@users.noreply.github.com>
2026-10-16 21:10:12 +00:00
Julian Knutsen
babf8d7ed7
perf(dolt): sweep the circuit-breaker dir once per process and age out closed files (#7593)
* perf(dolt): sweep the circuit-breaker dir once per process and age out closed files

CleanStaleCircuitBreakerFiles is documented as an init-time cleanup that
is "fast — typically 0-2 files", but it reads every breaker file in the
live directory on every server-mode store open. Two things made that the
dominant fixed cost of a bd write on a busy dev host:

- Every host:port:db a process opens leaves a closed-state file behind,
  and ephemeral test-server ports mint a new name each time. Closed files
  in the live directory were never removed, so one host's
  $TMPDIR/beads-circuit held 2335 of them.
- A write command opens its store twice — the non-mutating preview open
  for the workspace-identity check (be-0gfcs), then the writable open —
  and each open ran the full sweep. One `bd update` opened 4755 breaker
  files.

The store-open path now sweeps at most once per process
(cleanStaleCircuitBreakerFilesOnce); bd doctor --fix still calls the
sweep directly. And the live directory ages out closed files untouched
for 24 h exactly as the legacy /tmp/beads-circuit sweep already did
(legacyClosedSweepTTL becomes closedSweepTTL for both). A closed file
that old is indistinguishable from no file: readState reads an absent
file as closed, and a closed state's failure count only matters inside
circuitFailureWindow. A breaker in use is rewritten on every successful
open, so its mtime stays fresh and it is kept.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(dolt): pin closedSweepTTL's 24h width; document what the sweep removes

Maintainer fixup for the PR #7593 review (iteration 1; all non-gating):

- TestCleanStaleCircuitBreakerFiles_AgesOutClosedLiveFiles stamps its closed
  files at literal ages instead of ones derived from closedSweepTTL: a file
  idle 25h must go and one idle 23h must survive. A TTL lowered to
  5m/61s/30s/22h or raised to 26h/48h, or a use site compared against
  circuitStaleTTL, now fails the test; every one of them passed before.
- CleanStaleCircuitBreakerFiles' godoc lists the closed-file age-out and the
  corrupt-file removal it already does, which directories it sweeps, and that
  store open reaches it once per process while bd doctor --fix calls it
  directly.
- The cleanServerCircuitState seam notes that a test driving the real sweep
  through it must start a fresh staleCircuitSweep.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-16 17:05:11 +00:00
Julian Knutsen
3862b03ff5
feat(http): gate await members on create and update over bd serve (#7589)
* feat(issueops): gate members on create and update, one bounds validator

The library half of serving gate await fields. Every local leg already
stores await_type, await_id, timeout and waiters on create, and await_id
on update; this adds the two missing pieces.

- IssuePatch.Waiters: a whole-value replacement of the waiter list (nil
  or empty clears, stored as []), mapped by UpdateFields and by the
  unit-of-work updateSpec, deep-copied by CloneUpdateRequest, and
  classified as a non-coordination signal on the dolt leg. Duplicates are
  stored as given; dedup stays the CLI's.
- ValidateGateFields / ValidateGatePatch with *issueops.GateFieldError
  {Field, Detail} wrapping ErrValidation, mirroring ReclaimFieldError so a
  front door can name the member. The rules are the columns' only:
  await_type <= 32 characters, await_id <= 255 characters, the JSON
  waiter list <= one TEXT column (65535 bytes), and timeout >= 0. No
  await_type allowlist, no gate-type-only rule, no timer-needs-timeout
  rule. Run from ValidatePublicCreateRequest (every create shape on every
  backend) and ValidateUpdateRequest (await_id, waiters).
- Conformance: LifecycleCreate WritesTheGateMembers / BoundsTheGateMembers,
  LifecycleUpdate PatchesTheGateMembers / BoundsTheGateMembers and
  BatchApply CarriesTheGateMembers, each bound asserted from both sides,
  wired on the dolt, embedded and uow legs and parked on the http leg on
  the live refusal rows until the wire slices land.
- httpclient: IssuePatch.Waiters is refused before the dial on both patch
  documents under a new W-IssuePatch.Waiters row (refuse-not-drop), until
  the wire publishes it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(httpapi): publish the gate members on create and update

The wire half of serving gate await fields, over the library slice.

- createIssue and the issues:batchApply create item publish await_type,
  await_id, timeout (nanoseconds, the encoding Issue.timeout reads back)
  and waiters, behind the new behavior token issues.create.gates (one token
  for both create shapes, the issues.create.defaultPriority precedent).
- updateIssue's patch and the batchApply update item's patch publish
  await_id and waiters (whole-value replacement; [] clears), behind
  issues.update.gates. await_type and timeout are not patch members: no
  local update route changes them. Neither is nullable; null is refused
  naming the member, as on title.
- The handlers only decode and project (createGateMembers,
  patchGateMembers). The wire is no stricter than the library: every bound
  is the role's *issueops.GateFieldError (await_type <= 32, await_id <=
  255, the JSON waiter list <= one TEXT column, timeout >= 0), mapped to the
  existing 400 invalid_argument with param naming the member at its level
  (await_type, patch.waiters, items[i].create.timeout,
  items[i].update.patch.await_id). The spec mirrors exactly that:
  maxLength 32/255 and minimum 0, and no minLength/maxLength/maxItems on
  waiters. Blank and duplicate waiters and any await_type reach the role
  untouched.
- Spec: the four schemas, the capability prose, and the stale sentences
  (createIssue's "await_* publishes on no operation, read or write" and the
  patch's "spec_id and await_id publish nowhere yet"). apigen regenerated
  (from the remote drift test's diff), wireshape golden regenerated:
  additive entries and sides only, wire_revision stays 2.
- Client: the two tokens are mirrored in wire.behaviorCapabilities so the
  vocabulary stays set-equal with the server's. The client still refuses
  every gate member before the dial (the client slice flips that); the write-ledger
  pending sets name the newly published members under their live refusal
  rows, and the W-CreateRequest.Issue / W-IssuePatch.AwaitID /
  W-IssuePatch.Waiters reasons, the served-leg park reasons and the S4b
  waiver now say the server publishes them and this client does not send
  them yet.

Tests: internal/httpapi/gate_fields_test.go pins member mapping on all four
bodies, clear-by-empty on both patches, null and unpublished-member refusal,
GateFieldError -> param at every nesting level, and both tokens advertised.
Mutating the handlers (dropping the mapping, the GateFieldError arms or the
tokens) turns them red.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(httpclient): send the gate members behind their tokens

The client half of serving gate await fields, over the wire slice.

- Create: await_type, await_id, timeout (nanoseconds) and waiters join the
  create vocabulary on createIssue and batchApply create items (now
  twenty-four members), sent verbatim when populated. Patch: await_id (by
  presence; "" clears) and waiters (whole-value; nil/empty travels as [])
  are encoded on both updateIssue's and the batchApply update item's patch.
- Token gate before the dial: a create naming any gate member needs
  issues.create.gates, a patch naming one needs issues.update.gates. Without
  the token the request refuses with wire.NewCapabilityError naming it
  (typed ErrUnsupported{Capability}), never a dropped member. batchApply
  checks the whole plan once, so no partial plan is sent. A request naming
  no gate member never reads either token. Store.requireCapabilityBeforeDial
  is the shared helper; requirePersistenceMove now uses it.
- Bounds read back typed: the server's 400 invalid_argument/invalid_value
  whose param names a gate member at any level (await_type, patch.waiters,
  items[i].create.timeout, items[i].update.patch.await_id) is rebuilt as the
  role's *issueops.GateFieldError, the problem kept in the chain (and a batch
  refusal still becomes its *ItemError). Without this the served bounds
  contracts could not pass.
- Ledger: W-IssuePatch.AwaitID RETIRED; the library slice's W-IssuePatch.Waiters row
  deleted (it never shipped outside this branch); W-CreateRequest.Issue and
  W-CreateItem.Issue no longer cover the gate members; new
  W-CreateRequest.Gates (refuse, Capability issues.create.gates) records the
  token-gated late refusal for both create shapes. engdocs ledger
  regenerated from Ledger() (262 rows: 201 refuse, 34 degrade, 27 retired).
- Write-ledger table: the wire slice's pending sets are flipped into carried.
- Served leg: the five gate cases (LifecycleCreate Writes/Bounds,
  LifecycleUpdate Patches/Bounds, BatchApply Carries) are un-parked and run;
  the S4b waiver and served comments now say await_id is sent.

Tests: internal/httpclient/gate_fields_test.go (encode on all four bodies,
clears, old-server refusal per member naming the token with no dial, no
partial batchApply plan, plain requests unaffected on old and new servers,
GateFieldError rebuilt from the problem). Disabling the three pre-dial
checks turns the refusal tests red.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* build(bazel): auto-sync BUILD files

Applied from the bazel-sync-patch artifact of https://github.com/gastownhall/beads/actions/runs/38841144413
(`make bazel-sync`: gazelle, tools/bazel/go_srcs.py, bazel mod tidy).
Only BUILD.bazel, MODULE.bazel and MODULE.bazel.lock are ever pushed; see
scripts/bazel-autofix-push.sh.

* docs(issueops): say the gate timeout rule is no column's

The gate validator's doc said its rules were the stored columns' and that
no command line can produce a negative timeout. The widths are the
columns' (VARCHAR(32), VARCHAR(255), TEXT), but timeout_ns is a BIGINT,
so a non-negative timeout is a rule no column enforces; the wire
publishes it as minimum 0. The raw store creates (CreateIssue,
CreateIssues) and import do not run the validator, so bd gate create
--timeout=-5m, a formula step's gate.timeout and a template clone still
store a negative timeout until they move onto the role.

Also say that GateFieldError.Detail is prose whose HTTP copy ends with
the server's own clause, so a caller dispatches on Field, and why a
create omits an empty waiter list while a patch sends [] to clear it.

Comments only; no behaviour change.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(corpus): gh:run await_id persist refuses by the missing gate token

#7604 (now on main) made bd gate check persist a gh:run gate's discovered
run ID through Lifecycle.Update. Its served test expected the HTTP
client's W-IssuePatch.AwaitID refusal, which persistGateCheckAwaitID
rewords to name await_id. This PR retires that row: the patch carries
await_id behind issues.update.gates. Through the reference proxy, whose
server advertises no gate token, the persist is now refused before the
dial by the capability error ("bd serve at ... does not advertise
capability "issues.update.gates", which updateIssue requires"), so the
merged tree failed the subtest's await_id and updateIssue assertions.

The subtest now expects that refusal: no writes, the error names the
gate, the missing token and the server, no role or store method leaks,
and await_id is untouched. persistGateCheckAwaitID already returns a
capability error as the role raised it (#7604 pins that), so no code
changes. Against a server that advertises the token the persist goes
through: a scratch probe against the in-process server recorded
await_id=4242 and closed the gate.

Also correct the gate.go comment and the CHANGELOG entry that still said
the HTTP client refuses every await_id patch.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <adopt-pr@gascity.com>
2026-10-16 17:02:27 +00:00
gascity-eddie[bot]
8d2b4152f7
fix: commit server-mode config and memory writes (GH#4078, carries #4383) (follow-up to #7299) (#7608)
* feat(storage): add CommitConfigOnly — scoped config-table commit (GH#4078)

Generic Commit() excludes the config table by design (GH#2455) and
CommitWithConfig stages everything via '-Am' — neither can commit an
intentional config write without sweeping concurrent operations' dirty
tables in server mode. CommitConfigOnly stages ONLY config (via the
existing doltAddAndCommit helper) and commits it.

- VersionControl interface method with GH#4078/GH#2455 rationale
- DoltStore: scoped DOLT_ADD('config') + DOLT_COMMIT
- EmbeddedDoltStore: aliases Commit — single-process, so '-Am' cannot
  sweep concurrent work (same rationale as its CommitWithConfig alias)
- Tests against a real dolt-sql-server container: scoped commit lands,
  concurrent dirty issues row survives uncommitted, and generic
  Commit()'s config exclusion is pinned as a regression guard

(cherry picked from commit 61ba8f05c9)
Carry note: rebased onto b1bee123b. The two version_control.go conflicts
were additive (upstream added CommitMergeResolution/CommitAll in the same
spot); resolved by keeping both sides, no change to this commit's code.

* feat(storage): CommitConfigUserKVOnly — screen memory commits to kv.* rows (GH#4078)

Carry reconciliation for the #4383 review: CommitConfigOnly stages the whole
config table, so a memory write's commit in server mode could sweep a
concurrent writer's half-applied internal key (issue_prefix above all), the
row-level form of the GH#2455 hazard. The review asked for #4412's kv.*
screen at the remember/forget sites.

CommitConfigUserKVOnly applies that screen before the scoped commit: if any
dirty config row is outside kv.*, it commits nothing and names the keys,
pointing at `bd dolt commit` (CommitAll), the same remedy the pre-pull
refusal gives. User kv.* rows already dirty are committed with the memory,
as the pre-pull auto-commit would. The screen query moves out of
assertDirtyConfigUserKVOnly into dirtyInternalConfigKeys so both paths share
it; the pre-pull refusal text is unchanged. Embedded aliases Commit, as its
CommitConfigOnly does.

Also softens the CommitConfigOnly interface doc (review nit): embedded mode
commits the whole working set, not only config.

Tests (server container): user kv.* rows commit; a dirty internal key is
refused with HEAD and the working set untouched.

* fix(storage): stage config's lookup-table projections with CommitConfigOnly (GH#4078)

Carry reconciliation for a change on main since #4383 was written:
SetConfig now writes status.custom into custom_statuses and types.custom into
custom_types in the same transaction as the config row (SyncConfigTables),
and reads of those sets are table-first. CommitConfigOnly staged config
alone, so `bd config set status.custom ...` in server mode would commit the
row and leave its projection dirty: HEAD would carry the new string with the
old lookup set, the split the projection exists to prevent. The
transactional SetConfig already marks the synced table dirty for this reason.

CommitConfigOnly now stages config plus issueops.ConfigProjectionTables(),
a list kept next to SyncConfigTables. Clean tables stage nothing, and the
projections are written only from config, so no other kind of work is
swept. CommitConfigUserKVOnly keeps staging config alone: memory rows have
no projection.

Test (server container): status.custom and types.custom writes leave
config, custom_statuses and custom_types all clean after CommitConfigOnly.

* fix(cli): commit config writes in server mode — remember/forget/config set (GH#4078)

In server mode nothing ever commits config-table writes: maybeAutoCommit
skips SQL-server modes entirely, and the explicit flush commands call
Commit(), which excludes config (GH#2455) and silently no-ops on a
config-only working set. Every bd remember / bd config set then strands
the working set dirty and the next 'bd dolt pull' fails with 'cannot
merge with uncommitted changes' until a manual raw-SQL flush.

commitConfigWrite commits the config table immediately after an
intentional config write, scoped via CommitConfigOnly so concurrent
operations' dirty tables are never swept. Server modes only — embedded
'-Am' auto-commit and the unflagged-writes sweep already cover config.
This restores the v1.0.1 always-commit behavior from PR #3052 that the
unified-PostRun refactor regressed.

Wired sites: bd remember, bd forget, bd config set, bd config unset,
bd config set-many (one commit for the whole batch). Commit failures
surface loudly per each site's error convention.

(cherry picked from commit 64f77a1963)
Carry note: rebased onto b1bee123b. config set/unset now write through the
WorkspaceConfig role and remember/forget through the memoryops role, so the
conflict resolution re-anchors each commitConfigWrite call right after that
site's noteDirectConfigWrite/noteDirectMemoryWrite (set-many: after its
direct-route store.SetConfig loop) and uses each RunE's current error
convention (return HandleError / HandleErrorRespectJSON instead of os.Exit).
No change to commitConfigWrite or its tests in this commit.

* fix(cli): reconcile commitConfigWrite with current main (GH#4078)

Carry reconciliation on top of the cherry-picked commit, for two things
that changed on main since #4383 was written:

- Proxied route: proxied mode never opens a store, and its WorkspaceConfig
  and memory roles commit each write inside their own unit of work
  (noteDirectConfigWrite / noteDirectMemoryWrite say so). commitConfigWrite
  now returns early on usesProxiedServer() instead of relying on the nil
  store, and the proxied test asserts the no-op.
- dolt.auto-commit: bd-4wamg made batch/off real deferral in server mode,
  and #5830 made `bd dolt commit` commit config (CommitAll). A config write
  under batch/off now defers with every other server-mode write instead of
  committing regardless; mode on (the default) commits as before. This
  matches embedded mode, where the PersistentPostRun auto-commit honors the
  same setting.

The doc comment drops the stale "-Am auto-commit and unflagged-writes
sweep" wording and says why rename-prefix and migrate are not wired.

* fix(cli): commit remember/forget through the kv.*-screened variant (GH#4078)

Carry reconciliation for the #4383 review: bd remember and bd forget write
user kv.* data, so their server-mode commit now goes through
CommitConfigUserKVOnly via a new commitMemoryWrite helper. A dirty internal
config key from another writer is refused with the keys named and the
`bd dolt commit` remedy, instead of being swept into the memory's commit.
bd config set/unset/set-many keep CommitConfigOnly: the operator is writing
config deliberately there, the same trust level as `bd dolt commit`.

commitConfigWrite and commitMemoryWrite share one guard path
(commitConfigTable), so both no-op on embedded, proxied, and batch/off.

* docs: changelog entries for the GH#4078 config-commit fixes

Server-mode config writes (bd remember/forget, config set/unset/set-many)
now commit immediately via the scoped CommitConfigOnly, and the explicit
'bd dolt commit'/'bd vc commit' are config-inclusive and truthful.

(cherry picked from commit c412b96b96)
Carry note: rebased onto b1bee123b. The hunk's old context now sits in a
released section, so it is moved under [Unreleased] > Fixed. The second
entry (bd dolt commit / bd vc commit) is dropped: that fix landed separately
as #5830 (CommitAll) and is already in the changelog.

* docs(changelog): update the GH#4078 entry for the carried fix

Rewords the cherry-picked entry for what the fix does on today's main:
batch/off defer config writes like every other server-mode write, the
commit stages config's lookup-table projections, remember/forget are
screened to kv.*, and #4412's pre-pull commit now absorbs dirty kv.* rows,
so the pull-side symptom is a refusal on internal keys rather than the
merge error. Drops the #2455 link, which does not resolve in this repo
(the number is a source-comment reference).

* test(cli): pin server-mode config writes leave no config dirt (GH#4078)

User-level regression test for the carried fix, on the direct SQL-server
route against the suite's Dolt server: bd remember, bd config set
(status.custom), bd config set-many (types.custom + status.custom), bd
config unset and bd forget must each leave config, custom_statuses and
custom_types clean in dolt_status. A final step dirties an internal config
key out of band and checks that bd remember refuses, names the key, and
leaves config dirty for bd dolt commit.

On b1bee123b (without the fix) it fails at its first step:
  after bd remember --key gh4078 server-mode memory: config left
  uncommitted in the working set (GH#4078)
With memory writes routed through the unscreened commit, it fails at the
last step: bd remember committed over a dirty internal config key.

* docs(cli): say why config writes skip the kv.* screen (GH#4078)

commitConfigWrite stages every dirty config row, while commitMemoryWrite
refuses internal keys. Record the reason in the code, not only in the
commit history: bd config set/unset/set-many are deliberate config
writes, the same trust level as bd dolt commit.

* build(bazel): list the GH#4078 test files in BUILD srcs

The rows the Bazel / test job's 'BUILD files in sync' check printed at ec27abd1d. Without them the Bazel tiers never run config_commit_write_test.go, config_commit_server_integration_test.go or commit_config_test.go, and the server-Dolt storage tier fails its shard check for the six storage tests.

* fix(storage): say the kv.* write is already saved when the screen refuses (GH#4078)

When CommitConfigUserKVOnly refuses, the memory row's SQL transaction has already committed, so exit 1 read as 'not stored'. The refusal now says the write is in the working set and that bd dolt commit publishes it with those keys.

* fix(httpclient): add the scoped config commits to the no-op commit family

The PR adds CommitConfigOnly and CommitConfigUserKVOnly to
storage.VersionControl. Since the PR's merge base, main added
internal/httpclient, whose Store asserts storage.DoltStorage, so the
tree merged with main did not build (*Store does not implement
storage.DoltStorage: missing method CommitConfigOnly).

Add both methods to httpclient's commit family as no-ops, for that
family's D3 reason: a write over this client is durable once the server
answers, so there is nothing to flush, and refusing would fail a command
whose write already landed. Regenerate unsupported_gen.go: only its
trailer count moves (201 -> 203 methods), because both methods are
hand-written rather than stubbed. Extend the two tests that pin the
family by name, TestCommitFamilyIsNoOp and
TestCommitFamilyIsNotOnTheUnsupportedMap, to cover both methods.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Eddie the Engineer
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: sms <114885497+seanmartinsmith@users.noreply.github.com>

* fix(config): screen the staged set, warn on refused memory commits (GH#4078)

Review fixes for the adopted PR, iterations 1 and 2.

CommitConfigUserKVOnly screened the working set on one pooled
connection and staged config on another, so an internal key written in
between (issue_prefix above all) was committed with a memory: the
GH#2455 hazard the screen exists to refuse. The screen, DOLT_ADD and
DOLT_COMMIT now share one pinned connection (stageAndCommit), and the
screen runs twice: on the working set before DOLT_ADD, and on what
DOLT_ADD actually staged after it, because sessions on one branch share
the working set. A refusal after staging unstages config again with
DOLT_RESET, so the refused rows are not left staged for the next commit
on the branch. Two residuals are documented at the code: another commit
path's own DOLT_ADD of config landing between the second screen and
DOLT_COMMIT is absorbed, as for every scoped commit, and the table-wide
reset can unstage such a commit's config staging. The refusal is now a
*storage.ConfigCommitRefusedError carrying the keys.

The two config-table commits also stop widening their staging list
with the versioned-history tables: a config write mints no version
rows, and that list's issues table swept concurrent issue work into a
config or memory commit.

bd remember and bd forget exited 1 after their write had landed
whenever a non-kv.* config key was dirty, including keys bd's own
unwired writers leave, such as the issue_prefix of bd rename-prefix.
They now warn on stderr with the keys named and exit 0, as v1.0.1
(#3052) did on a failed memory commit. Under --json they add
committed: false, commit_error and dirty_config_keys, only when the
commit did not happen, so a committed write's payload keeps its shape.
The CHANGELOG entry now says so instead of claiming that the refusal
restores the v1.0.1 behavior.

bd config unset of a yaml-only secret key deleted the credential row an
older bd had stored, but never committed the delete, so HEAD, which
bd dolt push publishes, kept the credential while --json reported
database_row_removed. The delete is now committed like any other config
write. A failed commit is a warning, and database_commit_failed under
--json, because the delete and the YAML unset both landed.

A commit error the store marks storage.ErrCommitIndeterminate, such as
a connection lost during DOLT_COMMIT, may have landed, so neither
surface reports it as a definite non-commit any more. --json adds
commit_indeterminate (remember, forget) or
database_commit_indeterminate (the secret-row unset) beside
committed: false and database_commit_failed, which stay, so a consumer
that reads only those is still told to commit; the secret-row warning
and the note a failed YAML unset carries say the removal may not be
committed and point at bd dolt commit.

commitConfigWrite's doc now says why a failed commit fails the config
verbs while the memory verbs only warn, and names bd kv set/clear and
the kv.memory rows bd import writes as config writes it does not wire
yet. CommitWithConfig's interface doc drops bd config set, which on
the SQL-server route now commits through CommitConfigOnly, and points a
command committing only its own config write at the scoped commits.

Tests: a live-store test writes an internal key between the screen and
DOLT_ADD and requires a refusal that moves no HEAD and leaves nothing
staged; a sqlmock test pins the refusal SQL, including a failed
unstage; a versioned-history-on test requires a dirty issues table to
stay dirty through both scoped commits; cmd/bd fakes pin the
warn-and-succeed contract, the JSON fields, the secret-row commit, and
the indeterminate qualifiers in both payloads and both texts; and the
server E2E gains a secret-row step plus a bd rename-prefix step ahead
of bd remember and bd forget.

Pre-approval CI repair: tools/nogo/unvalidated_integration_tests.txt
now lists the PR's integration-tagged
cmd/bd/config_commit_server_integration_test.go. Since #7486 no nogo
lane analyzes integration-tagged test files, and
TestUnvalidatedIntegrationTestsFileIsExact requires the list to name
every one of them, so the new file failed //scripts:nogo_integration_scan_test
in the Bazel test and integration lanes.

Left for follow-ups, both pre-existing and not changed here: the
review's projection minor (config delete paths do not reconcile
custom_statuses/custom_types), and committing bd kv set/clear and
bd import's kv.memory rows on the SQL-server route.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: sms <114885497+seanmartinsmith@users.noreply.github.com>

---------

Co-authored-by: Aaron Lippold <lippold@gmail.com>
Co-authored-by: sms <114885497+seanmartinsmith@users.noreply.github.com>
Co-authored-by: Eddie the Engineer <adopt-pr@gascity.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
2026-10-16 13:48:30 +00:00
gascity-eddie[bot]
cfadcde8f4
fix(federation): treat missing peer row as credential-free remote (follow-up to #5128) (#7606)
* fix(federation): treat missing peer row as credential-free remote

federation status/fetch failed with "failed to get peer credentials:
not found: federation peer origin" when the Dolt remote existed but no
federation_peers credential row was stored (add-peer without --user, or
sync.remote bootstrap). That made status report Reachable=false while
bd dolt pull/push succeeded (GH#4837).

withPeerCredentials now treats storage.ErrNotFound as empty credentials
so normal Dolt credential discovery applies. Decryption and other errors
remain fatal.

Fixes #4837

* test(federation): cover withPeerCredentials missing-row path; expect Dolt unknown-remote error

TestFilteredPushUsesDedicatedLongTimeoutConnection asserted the pre-fix
"failed to get peer credentials" error for "nonexistent-peer". Under this
PR a name with no federation_peers row is a credential-free remote, so the
push proceeds to Dolt and fails there with "remote 'nonexistent-peer' not
found". Assert that instead; waitForPoolWait and
assertFederationStagingBranchAbsent are unchanged.

Add TestWithPeerCredentialsMissingPeerRow (server mode, skipIfNoServer):
withPeerCredentials on a name with no row invokes fn exactly once with nil
credentials and returns fn's result; a peer stored with credentials still
delivers them to fn. Deleting the new branch in withPeerCredentials now
fails the test.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* test(federation): build credentials_peer_test.go under Bazel; e2e row-less remote test (#5128 review)

Maintainer fixup on top of the contributor's two commits, from the
adopt-pr review of #5128 (iterations 1 and 2).

- internal/storage/dolt/BUILD.bazel: list credentials_peer_test.go in the
  dolt_test srcs. The target lists its srcs explicitly, so the new file
  was never compiled into the Bazel test binary, and the required "BUILD
  files in sync" step (make bazel-sync, then git status) would red
  BAZEL_TEST and the CI Gate. The step body now reports no drift.

- TestFederationVerbsWithoutPeerRow: the end-to-end case #4837 reported.
  The remote is registered with AddRemote alone (no federation_peers row);
  PushTo, Fetch and PullFrom must succeed against it, Fetch must move
  remotes/peer/main to the peer's HEAD, the pull must merge the peer's
  issue, and no peer row may appear. With the fix reverted it fails at
  PushTo with "failed to get peer credentials: not found: federation peer
  peer"; with Fetch made a no-op it fails at the tracking-ref check.

- withPeerCredentials doc comment: record the mistyped --peer trade-off
  (the name reaches Dolt, which rejects it), and that a name check added
  in front must consult PersistedRemoteInfos as well as ListRemotes so
  the GH#2118 cold-start window is not rejected.

- CHANGELOG [Unreleased] Fixed entry for #4837. It names the bd
  federation verbs only: bd dolt push/pull --remote go through
  PushRemote/PullRemote, which never look up a peer row, so they were
  not affected.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Vishnu Jayavel <vishnuj81093@gmail.com>

---------

Co-authored-by: Vishnu Jayavel <vishnuj81093@gmail.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
2026-10-16 12:07:30 +00:00
Alex
fefcdf6533
fix(delete): resolve ids exactly, never by prefix abbreviation (#7375)
* test(delete): pin that a stale id never deletes the issue it prefixes (be-ksd)

bd delete resolves each id through the abbreviation-tolerant resolver
(resolveAndGetIssueForMutation -> utils.ResolvePartialID). When the exact
id is absent, that resolver falls through to leading-prefix abbreviation,
so an id that was already deleted resolves to a different issue whose id
it happens to begin, and --force deletes that issue.

The shape is ordinary: hash length is adaptive (3 to 8 characters) so a
grown store holds short ids beside longer ones, --id makes any shape, and
a hierarchical child's id always extends its parent's. On a fixture, after
fx-t00003 was deleted, bd delete fx-t00003 resolved to fx-t000031; a
re-run of a finished --from-file list with --force --quiet deleted two
issues it never named, silently.

The test drives deleteCmd in-process against a stub DoltStorage (no cgo,
no Dolt server) and records every request handed to the Deleter role. It
covers the single-id --force path, the preview, a deleted parent's id with
an orphaned child, the --from-file batch path, and an abbreviation of one
and of two live issues; each refusal must say delete needs the full id
and name the ids the input matches. An over-refusal guard checks that
exact ids (full, bare hash, child, batch) still delete exactly what they
name. The six refusal cases fail at this commit; the guard passes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Signature: claude-code-claude-opus-5-5 on behalf of A3Ackerman

* fix(delete): resolve ids exactly, never by prefix abbreviation (be-ksd)

bd delete's direct route resolved each id with the abbreviation-tolerant
resolver, so an id that no longer exists resolved to the surviving issue
whose id it begins, and --force deleted that issue. Both of its resolution
sites now use resolveAndGetIssueForMutationExact (#5393's exact sibling,
until now used only by bd comment): the inline single-id path and
deleteBatch, which --from-file, multiple ids and --cascade take.

The resolver cannot tell a stale full id from an abbreviation. Hash length
is adaptive with store size, so one store holds 3- to 8-character ids side
by side and "fx-t00003" is both a valid full id and a valid abbreviation
of "fx-t000031". The rule therefore has to be per command, the way #5393
made it for bd comment. Delete is exact on both preview and --force: the
preview prints the next command to run ("To proceed, run: bd delete
fx-t000031 --force"), so a tolerant preview hands the caller the wrong
deletion to confirm.

This matches the team-server route, which passes ids to issueops.Deleter
unresolved, and DeleteRequest.IDs' own contract ("IDs ARE EXACT"). The
partial-ID notice (#6302) did not protect this path: --quiet silences it,
and a --force --quiet re-run of a finished --from-file list deleted two
issues it never named.

The refusal says what to do: "bd delete needs the full issue id: no issue
has the id "t000", which is the start of fx-t000031". To name the matches,
ResolvePartialIDExact now returns *utils.AbbreviatedIDError, which carries
the input and the sorted matches, wraps ErrAbbreviatedIDNotAllowed and
keeps the old error text, so bd comment's errors.Is checks and messages
are unchanged.

Cost: bd delete <abbreviation> is refused, naming the full ids, so the
caller retypes one. Exact ids, bare full hashes, child, cross-prefix and
routed ids resolve as before. The other deleteBatch callers (cleanup, wisp
gc, mol burn) pass full ids read from the store, so nothing changes for
them except under a race, where they now refuse instead of reaching a
neighbor. Reads and other writes keep abbreviation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Signature: claude-code-claude-opus-5-5 on behalf of A3Ackerman

* docs(ids): say bd delete requires the full id, like bd comment (be-ksd)

The Working with IDs section named bd comment as the one command that
refuses abbreviations and said every other mutation keeps them. bd delete
now requires the exact id too, so the section covers both, says why a
delete reaches the wrong issue so easily (an already-deleted id is still
the start of its child's id, or of a longer hash), and adds the bd delete
lines to the example block. Best Practices #1 names both commands.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Signature: claude-code-claude-opus-5-5 on behalf of A3Ackerman

* docs(changelog): link the bd delete exact-id entry to #7375 (be-ksd)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Signature: claude-code-claude-opus-5-5 on behalf of A3Ackerman

* fix(delete): report a gone parent's id as a lost race under --if-revision (be-ksd)

Review-loop iteration 1 for #7375: a maintainer fixup on top of the
contributor's branch.

Major (codex+claude): a guarded `bd delete --if-revision` whose target
was deleted, and whose id is still the start of a surviving id (a gone
parent with a live <parent>.1 child), exited 1 with the abbreviation
refusal. Any other gone target gets the documented lost race
(precondition_failed, exit 13), and gascity's conditional deletes
switch on that code, so a benign race read as a hard failure.

- reportIfRevisionTargetGone (the single-id path) also takes
  utils.ErrAbbreviatedIDNotAllowed as a gone target: under delete's
  exact resolution that id names no row either.
- The refusal is a typed error, deleteAbbreviationError, that unwraps
  to storage.ErrNotFound beside the *utils.AbbreviatedIDError, so the
  --cascade caller of deleteBatch classifies it through
  reportIfRevisionFailure the same way. errors.New(msg) dropped every
  sentinel.

Minor (claude), taken: the refusal leads with the not-found fact,
`issue "t000" not found: bd delete needs the full issue id, and "t000"
is only the start of fx-t000031`, so gascity's isBdNotFound reads an
already-gone full id that begins a survivor's id as the idempotent
not-found it reads for any other gone id. The CHANGELOG entry quotes
the new text and the --if-revision outcome.

Minor (gemini), declined: threading the source command into the
refusal's wording. deleteBatch's other callers (bd cleanup, wisp gc,
mol burn) pass store-minted full ids that no user retypes, and the text
now leads with "issue X not found", which holds for every caller.

Tests:
- TestDeleteIfRevisionGoneIDThatPrefixesAnotherIssue: single id,
  --json, --cascade and --cascade --json each exit 13 with "precondition
  failed: issue no longer exists" and the precondition_failed body,
  never the refusal text, and the guarded write never runs. Two
  no-survivor rows are the plain gone id they must match.
- TestReportIfRevisionTargetGone: abbreviation rows, guarded and not.
- TestDeleteRefusesStaleIDThatPrefixesAnotherIssue also asserts the
  refusal says "not found".
- TestEmbeddedIfRevisionDeletePreflightGoneParentIsPreconditionFailed,
  end to end on embedded Dolt: delete a parent, its child survives and
  `bd show <parent>` resolves to it; a guarded delete of the parent
  exits 13 on the single-id and --cascade paths and leaves the child.
  The TestEmbedded name is what puts it in a CI lane; its line in
  embedded-cmd-test-shards.txt is the generator's incremental --write.

Mutation checks, each reverted afterwards: dropping the
ErrAbbreviatedIDNotAllowed match fails the single-id cases and the
abbreviation_guarded row; dropping ErrNotFound from Unwrap fails the
--cascade cases; dropping "not found" from the text fails every
refusal case; reverting both fixes fails both embedded legs (exit 1,
the refusal text, no "precondition failed").

Validation: the focused delete and if-revision tests; the embedded pair
under BEADS_TEST_EMBEDDED_DOLT=1; the scripts shard-manifest policy
tests and the generator's --check; gofmt; git diff --check. The full
./cmd/bd and ./internal/utils packages (-short, BEADS_TEST_SKIP=dolt)
pass but for TestRunApplyAllOK, which fails only because this worktree
sets core.hooksPath=/dev/null ("mkdir /dev/null: not a directory") and
passes with core.hooksPath overridden.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
2026-10-16 11:25:54 +00:00
gascity-eddie[bot]
3a79fbad4c
test(resolver): pin leading-prefix partial-ID matching for gcy-g4o (follow-up to #4376) (#7605)
* fix(resolver): use last-segment prefix match to prevent wisp substring collision

When two beads share overlapping hash fragments (e.g. gcy-oqf and
gcy-wisp-goqfo), the previous strings.Contains check incorrectly matched
"goqfo" as containing "oqf" and silently returned the wrong bead.

The fix changes the fuzzy match from substring containment to a
last-dash-segment prefix comparison:
  lastSegment("wisp-goqfo") = "goqfo" → HasPrefix("goqfo","oqf") = false
  lastSegment("wisp-t3st")  = "t3st"  → HasPrefix("t3st","t3st") = true

This preserves the intended use case (resolving wisps by their trailing
segment, e.g. "t3st" → "bd-wisp-t3st") while eliminating false
substring hits.

Also adds BD_NO_FUZZY env var: when set, partial resolution is skipped
and only the exact-ID fast path runs. This lets callers opt out of all
fuzzy matching when they need strict lookups (e.g. to avoid cache-lag
ambiguity).

Fixes: gcy-g4o (fuzzy short-ID resolver silently redirects to wrong bead)

* test(resolver): pin prefix-only semantics for plain-hash partial match

* fix(resolver): restore leading-prefix ID matching, pin gcy-g4o cases

#4939 already made partial matching leading-prefix only (strings.HasPrefix
in place of strings.Contains at both call sites), which is what fixed the
gcy-g4o collision: gcy-oqf can no longer resolve to gcy-wisp-goqfo through
the substring "oqf".

The last-dash-segment matcher added on top of that compared only the text
after each id's last dash, so inputs naming no issue resolved to unrelated
beads: hr-8wnq to a local bd-8wnq2 (which also pre-empts its hr- prefix
route), gcy-wisp-g4o to gcy-g4oab, and a trailing dash to any hash. The
first loop also began claiming wisps as abbreviations, so an exact bare
wisp hash turned ambiguous beside a longer sibling. Restore main's matcher
and remove hashSegmentPrefixMatch and lastDashSegment.

Drop BD_NO_FUZZY and its test. The gate returned before prefix
normalization, so documented exact forms such as a bare hash failed, and
it skipped the ErrAbbreviatedIDNotAllowed sentinel that callers branch on.
Exact-only resolution already exists as ResolvePartialIDExact; a
user-facing strict mode is left as a follow-up.

Keep the gcy-g4o regression tests, reworked so they reach the partial
matcher:
- TestResolvePartialID_WispSubstringCollision stores only the wisp, so
  gcy-oqf and oqf must miss instead of hitting the exact fast path.
- TestResolvePartialID_WispLastSegment adds a longer wisp sibling, so the
  wisp fallback's exact branch is what resolves t3st.

Add Docker-free and routed coverage:
- TestResolvePartialID_FakeStore_QualifiedInputsNeverResolveToAnotherBead:
  hr-8wnq, gcy-wisp-g4o, bd-wisp-t3, abc, bd-ol-, aap- and - stay
  not-found on both resolvers, each beside a control abbreviation that
  resolves.
- TestResolvePartialIDMatchesLeadingPrefixOnly,
  TestResolvePartialIDSilentOnExactWispHash and
  TestResolvePartialIDNotifiesOnWispHashAbbreviation replace the
  helper-level tests.
- TestRoutedResolutionLooksPastLocalNearMiss: hr-8wnq reaches its hr-
  route despite a local bd-8wnq2, on the read and mutation paths.

With the last-segment matcher swapped back in, the table test (7 of 8
cases), the routed test, TestResolvePartialIDSilentOnExactWispHash and
TestResolvePartialID_WispLastSegment fail. With strings.Contains, the
table test (4 cases), TestResolvePartialIDMatchesLeadingPrefixOnly,
TestResolvePartialIDSilentOnExactWispHash and
TestResolvePartialID_WispSubstringCollision fail.

Co-authored-by: Saren <saren@dumstruck.com>
Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Saren <saren@dumstruck.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-16 11:25:14 +00:00
Julian Knutsen
fdec61870c
perf(schema): drop the per-open INFORMATION_SCHEMA.TABLES probes (builds on #7409) (#7596)
* perf(schema): probe per-open sentinel tables with SHOW TABLES, not INFORMATION_SCHEMA.TABLES

Every bd open runs cursorRealityFloor and the ignored-cursor gate, and both
asked INFORMATION_SCHEMA.TABLES whether a table exists: the sentinel tables
(wisps, wisp_dependencies) and the untrack scratch table. Dolt does not push
the TABLES predicate down any more than the COLUMNS one, so each probe walks
the catalog of every database on the server.

Measured 2026-10-08 on a shared dolt sql-server (dolt 2.1.10, 9 databases,
about 12 to 17 bd opens a second, on a v1.3.1 line that also carries three
floored sentinel tables): the server ran at 365 per cent CPU and these
existence probes were the largest single class of running statements. A
debug query log of one `bd list`, `bd ready` and `bd show` against a warm
proxied server shows 10 INFORMATION_SCHEMA statements per open before; with
this change and the sentinel-column change (#7297), 3. The statement count
per command is unchanged.

showTableExists reads SHOW TABLES LIKE with an exact name compare ('_' is a
LIKE wildcard). SHOW TABLES errors when no database is selected, so it is
used only where the session is already on the target database: the
sentinels run after the cursor probe has found the cursor table, and
readIgnoredCursorState runs after selectTargetDatabase or MigrateUp's
preparation. The cursor-existence probe in currentVersion keeps the
INFORMATION_SCHEMA form, because it must succeed on any session (be-bv7x).
Repair paths keep schemaTableExists; they run once per migration.

TestShowTableExistsMatchesInformationSchemaOnDolt checks the two forms agree
on a real Dolt for a present table, an absent one, a dolt_ignored
working-set table and a LIKE lookalike; the sqlmock helpers move to the SHOW
TABLES form. Stacked on #7297 (the sentinel-column change), which it depends on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Rebased onto #7592 (which carries #7297): main has since added the floored
sentinel tables (events, bd_events_journal, bd_events_seq) and #7592 the
leases carrier read, all through sentinelTableExists, so their ordered
sqlmock expectations move to expectShowTable too; lock_test.go conflicted
only on expectIgnoredSentinelProbes.

(cherry picked from commit ca0e921348)

* test(schema): pin the per-open table probes to INFORMATION_SCHEMA's answers

Red half of finishing the per-open INFORMATION_SCHEMA.TABLES removal on
top of #7409. Two gaps against the count the probes replace, both
measured on dolt 2.1.8 and 2.2.0:

- Case. INFORMATION_SCHEMA (like every table reference) matches table
  names case-insensitively, but Dolt's SHOW TABLES LIKE filter does not:
  a sentinel stored as BD_EVENTS_SEQ reads as absent, so the ignored
  cursor clamps to 21 and the tail replays on every open (its creators
  are INFORMATION_SCHEMA-guarded, so the replay never heals it).
- The cursor-existence probe in currentVersion is still an
  INFORMATION_SCHEMA.TABLES read, issued six times per bd update (~25-30
  ms each on a server hosting 28 databases). It must succeed on any
  session (be-bv7x), and SHOW TABLES fails with no database selected
  (1105) or on a session whose database was dropped (1049).

Tests:
- TestShowTableExistsAnswersLikeInformationSchema: the unfiltered listing,
  names compared case-insensitively; a '_' lookalike is not the table.
- TestCurrentVersionWithoutAUsableDatabaseIssuesNoFailingStatement: the
  schemata read answers 0 and nothing follows it.
- TestShowTableExistsMatchesInformationSchemaOnDolt (from #7409) now
  covers every per-open probe name under bd's own dolt_ignore patterns,
  a LIKE lookalike and a table stored in another case.
- Real sql-server: TestShowTablesProbes_SeeIgnoredTablesInAnyCase (a
  migrated store's cursors are believed on a fresh session, also with
  bd_events_seq stored as BD_EVENTS_SEQ) and
  TestCursorProbe_NoDatabaseSessionStaysUsable.
- The cursor-probe sqlmock helpers (schema, uow, cmd/bd) expect the
  schemata read and the SHOW TABLES listing.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* perf(schema): drop the last per-open INFORMATION_SCHEMA.TABLES reads

Finishes what #7409 starts. After it, every bd write still made six
INFORMATION_SCHEMA.TABLES point reads, all of them the cursor-existence
probe in currentVersion (5x schema_migrations, 1x
ignored_schema_migrations). Dolt does not push that predicate down: each
costs 25-30 ms on a sql-server hosting 28 databases, against ~0.6 ms for
SHOW TABLES or the schemata read (dolt 2.1.8; same on the shared server).

- currentVersion probes the cursor through tableInCurrentDatabase: an
  information_schema.schemata read for DATABASE(), then SHOW TABLES. The
  probe must succeed on any session (be-bv7x), and SHOW TABLES fails with
  no database selected (1105) or on a session whose database was dropped
  (1049); the schemata read answers 0 in both cases, which reads the
  cursor as absent exactly as the TABLES count with
  TABLE_SCHEMA = DATABASE() did.
- showTableExists reads the unfiltered SHOW TABLES listing and compares
  names case-insensitively. Dolt's LIKE filter on SHOW TABLES is
  case-sensitive while the INFORMATION_SCHEMA count (and every table
  reference) is not, so the LIKE form read a sentinel stored as
  BD_EVENTS_SEQ as absent and replayed the ignored tail on every open.
  Without a LIKE pattern the '_' wildcard concern goes away too; the
  listing is one round trip either way (34 rows, ~0.6 ms).

Verified equal to the INFORMATION_SCHEMA count on dolt 2.1.8 and 2.2.0
for every per-open probe name under bd's dolt_ignore patterns, on a
fresh session, after DOLT_CHECKOUT, on a revision database, with the
DSN database name in another case, with no database selected and on a
dropped database. The unit test gains a name-containing-the-table case
(events vs bd_events_journal).

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(schema): state the probes' preconditions; pin the 1049 cursor case

Review fixups for the per-open SHOW TABLES probes.

- readIgnoredCursorState and its alreadyConverged call site still
  described the scratch probe as an always-succeeding information_schema
  read bounded to one row. It is now an unfiltered SHOW TABLES listing,
  which fails with no database selected (1105) or on a dropped database
  (1049), so the doc states the precondition both callers meet: the
  session is already on the target database. The HEAD half's dolt_ignore
  read is named too; it runs only when the cursor is tracked at HEAD.
- tableInCurrentDatabase documents that its guard is check-then-act: a
  DROP DATABASE between the schemata read and SHOW TABLES still pins the
  session, and on a *sql.DB the two reads can land on different pooled
  connections, so every connection in the pool must be on one database.
- TestCursorProbe_DroppedDatabaseSessionStaysUsable is the 1049 half of
  the no-database server test: USE a database, drop it from another
  connection, then CurrentVersion must answer 0 without an error and the
  session must still see a database and table created afterwards.
  Probing the cursor with the bare SHOW TABLES prober fails it with
  "Error 1049 (HY000): database not found" (dolt 2.1.8).
- expectRemoteBackedSharedProbe uses `for range 2`, like its twin in
  migrate_consent_test.go.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Chris Sauer <csauer02@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-16 10:37:52 +00:00
gascity-eddie[bot]
9c5f897247
fix(storage): carry closed_by_session through the domain/UOW writer (#4662) (follow-up to #4798) (#7602)
* fix(storage): carry closed_by_session through the domain/UOW writer

Rebased onto current main, which changed what this PR needs to be.

The original diff added closed_by_session to the shared select/scan/
insert column sets (sqlbuild.IssueSelectColumns, issueops.ScanIssueFrom,
insertIssueIntoTable, issueUpsertColumns). All of that landed upstream
independently in #5191, so those hunks are dropped rather than
mechanically re-resolved against the lease-overlay refactor (0055) and
the row_lock/RowVersion hydration that arrived with it.

What is still missing on main is the second writer: db.insertIssueRow,
the domain/UOW statement behind proxied create, batch create and import,
whose column list contains no closed_by_session — not in the INSERT, not
in the ON DUPLICATE KEY UPDATE. The schema defaults the column to '' and
the now-shared scanner faithfully reads back whatever was stored, so a
proxied create silently discards the closing session while the same
issue created through the classic path keeps it. That is the GH#4662
data loss, still live, just relocated.

Add the column to insertIssueRow's insert, values and upsert
assignments, and note at the chokepoint why the two writers' column
lists have to stay in step.

Tests:
- Three new cases in the domain/db suite (real Dolt): insert -> get
  preserves closed_by_session on the issues table and on the wisps
  table, and the upsert rewrites it rather than keeping a stale value.
  All three fail against main and pass with the fix.
- The conformance case PromoteClosedWispPreservesSession from the
  original commit is kept: it runs on every backend and is the
  end-to-end GH#4662 regression (close a wisp with a session, promote,
  assert the session survives).

The nit about test comments asserting closed_by_session is absent from
IssueSelectColumns no longer applies — those comments are gone from
main, and the scan-parity fixture already counts 52 columns including
closed_by_session.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* fix(storage): name insertIssueRow's real callers and pin the upsert clear

Only Insert without CreateOnly reaches insertIssueRow: the CreateIssue
callers that leave CreateOnly unset, such as the quick, todo and gate
proxied servers and the create op of a proxied `bd batch` script.
Proxied create and the unit-of-work batch creator and applier set
CreateOnly, so they insert through the classic writer, and import and
promotion use the classic writer too. That writer and the shared scanner
have carried closed_by_session since #5191.

So the previous commit closes a latent gap in the second writer rather
than a loss on proxied create or import, and GH#4662's closed-wisp
promotion loss was already fixed by #5191. The conformance case is that
issue's regression test: it passes without the domain writer change,
which the domain/db Insert tests pin instead.

- Reword the insertIssueRow comment and the domain/db test docs to name
  the callers that reach the statement and its upsert half.
- Say in the PromoteClosedWispPreservesSession header that it guards the
  classic path only.
- Add UpsertClearsClosedBySession: rewriting a closed row with an open
  one must clear the stored session, not leave an open row with a stale
  closing session. It fails if the upsert keeps the stored value when
  the incoming one is empty.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Harry Miller <harry_miller@trimble.com>

---------

Co-authored-by: Harry Miller <harry_miller@trimble.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
2026-10-16 10:07:12 +00:00
Julian Knutsen
74a67c3c27
feat(prime): add prime.disabled so a session can turn off an inherited bd prime hook (#7595)
Claude Code merges a project's hook settings into every session started in that
checkout, including an orchestrator's agent sessions, which get their own
briefing. A higher-precedence settings file cannot remove a lower one's hook, so
such sessions run `bd prime` whether it helps them or not. The only per-session
channel their host controls is the environment.

prime.disabled (BD_PRIME_DISABLED) makes bd prime print nothing, the same as
outside a workspace (an empty envelope under --hook-json), whatever its other
flags. It returns before any workspace or store access. Like the prime memory
caps it is a yaml-only key, and it is validated with strconv.ParseBool, the set
that viper's GetBool reads.

Measured on a copy of a 751-memory store, bytes a hook injects:
- bd prime: 1,251,189 (the default is unchanged);
- BD_PRIME_DISABLED=true bd prime, --full, --memories-only: 0;
- BD_PRIME_DISABLED=true bd prime --hook-json: the empty envelope.

In the Gas City deployment that owns that store, 637 agent sessions in 48 hours
fired the inherited hook 1,271 times. Each firing injected a 2.3 KB preview after
host truncation. On top of that, the sessions read 1.81 MB of the persisted
output. That is about 4.8 MB of context these sessions do not use.

Tests:
- TestPrimeDisabledReadsTheConfigKey.
- TestPrimeDisabledPrintsNothingInAWorkspace. A staged workspace's briefing
  prints when the key is off; the test fails if the early return is removed.
- TestValidateYamlConfigValue_PrimeDisabled.
- The embedded TestEmbeddedPrime/prime_disabled_env and
  /prime_disabled_config subtests.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-16 08:31:08 +00:00
Julian Knutsen
de01b0f3e8
perf(schema): probe the column sentinel with SHOW COLUMNS, behind a carrier-table read (#7592)
* perf(schema): probe sentinel columns with SHOW COLUMNS, not INFORMATION_SCHEMA.COLUMNS

cursorRealityFloor runs on every store open. Its column sentinel used
INFORMATION_SCHEMA.COLUMNS, whose predicate Dolt does not push down: on a
sql-server hosting ~30 databases the probe scanned every database and took
5-6 s per open (measured 2026-09-17, dolt 2.3.x). A 15-store federated
gc ready spent ~80 s of its 90 s there. SHOW COLUMNS on the same server
answers in 0.15 s.

The existing content_hash prober is generalised into showColumnExists and
the open-path sentinel points at it; the repair paths keep
schemaColumnExists as is (they run once per migration, not per open). Mocks
pinned to the old form are moved to the SHOW COLUMNS form.

Rebased onto current main: lock_test.go conflicted only on context
(newer tests above expectIgnoredSentinelProbes); kept, helper comment
taken from this change.

(cherry picked from commit f02fff7bb5e1483cc8650afd20323d779db57334)

Claude-Session: https://claude.ai/code/session_01Ww5wAw2ybpByzduNHo5UYB
(cherry picked from commit e012f3a605)

* test(schema): pin the column sentinel probe to a never-failing carrier read

Red half of guarding #7297's SHOW COLUMNS sentinel probe (be-bv7x).
SHOW COLUMNS fails on a missing table, and on the store's multiStatements
DSN a failing statement pins the pooled Dolt session to its pre-statement
catalog: a table another connection creates afterwards stays "table not
found" on that session (reproduced on dolt 2.1.8 and on the 2.2.0 the
test harness pins). leases, the column sentinel's carrier table, is
exactly what every release-lineage store lacks.

- TestSentinelColumnProbeNeverAsksAMissingTable,
  TestSentinelColumnProbeReadsAPresentTableExactly and
  TestSentinelColumnCarrierProbeErrorIsNotTreatedAsContradicted pin the
  probe through the sentinelTableExists seam.
- TestIgnoredCursorReadWithLeasesAbsentIssuesNoFailingStatement walks the
  guarded cursor read with a strict ordered mock (currentVersion surfaces
  an unexpected statement; atLatest would swallow it).
- TestSentinelColumnProbe_PoolSeesLaterTables reproduces the pin on a real
  sql-server: a one-connection pool reads the ignored cursor with leases
  dropped, leases is re-created on another connection, and the pool must
  see it.

All of them fail against the unguarded probe. The ordered sentinel mocks
now expect the carrier read ahead of SHOW COLUMNS.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* perf(schema): read the column sentinel's carrier table before SHOW COLUMNS

#7297 moves the per-open column sentinel (leases.granted_node) from
INFORMATION_SCHEMA.COLUMNS, which Dolt does not push down, to SHOW
COLUMNS: on a sql-server hosting 28 databases that is 372 ms against
0.9 ms per write-capable open. But SHOW COLUMNS fails on a missing table,
and the carrier table is exactly what a release-lineage store lacks (no
leases before ignored/0012). On the store's multiStatements DSN a failing
statement pins the pooled Dolt session to its pre-statement catalog, so
the leases table the ignored replay then creates on another connection
stays "table not found" on that session (be-bv7x; dolt 2.1.8 and 2.2.0).

sentinelColumnExists now reads the carrier's existence through
sentinelTableExists, whose read succeeds whether or not the table is
there, and only sends SHOW COLUMNS to a table that exists. A missing
table still reads as a missing column, so the floors, the short-circuit
order and the decision in cursorRealityFloor are unchanged; a failed
existence read surfaces instead of reading as a contradiction.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Eric Cestari <eric@escapevelocity.fr>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-16 07:19:24 +00:00
Julian Knutsen
b303d494df
perf(http): answer a command's repeated reads from one response; bd show --json makes 1 request (port of bd-enterprise e626072b8) (#7588)
* perf(http): answer a command's repeated reads from one response (port of bd-enterprise e626072b8)

Over an http workspace, bd show <id> --json made three identical
GET /v0/beads/issues/<id> requests: the id resolver's exact probe,
routed.go's GetIssue and show's detail read all build the same URL. Text
show made six. Over a gateway each one is a sequential round trip, so the
request count is the command's latency floor (mc-5dkzp).

This ports bd-enterprise e626072b8 (bts-xpw.3.14.6, #217) verbatim, except
cmd/bd/main.go, where OSS still opens the identity preview inline, and
bd-enterprise's httpgw dialer test, which has no OSS counterpart:

- httpclient.WithSharedConnections: the identity-gate preview and the real
  open share one transport, so a write makes one handshake.
- httpclient.WithReadMemo and a wire.Client read memo, for one-shot
  commands only (not bd serve, --watch or --follow): identical getIssue,
  getSetting and listSettings reads are answered from the response already
  received. An entry ages from the server's answer (5 s), any write clears
  the memo, a read that overlapped a write is not stored, and expired
  entries are swept.
- bd list's three vocabulary reads come from one memoized listSettings.
- bd update --status / --if-status skips the custom-status read for a
  built-in status.
- The hook decorator skips its post-write re-reads when no executable hook
  exists for the event.
- bd dep add reads its confirmation titles before the write, from the memo.

Against a loopback bd serve: show --json goes from 3 requests to 1, text
show 6 to 3, update 5 to 3, close 6 to 4. Dolt and the other registered
backends receive the same ctx and ignore it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* feat(http): BD_HTTP_READ_MEMO=0 turns the per-command read memo off

An operator who suspects a stale read can rule the memo out without a
rebuild. Any other value, or none, leaves the default.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(http): pin bd show to one fetch per issue (mc-5dkzp)

TestE2E_PerCommandRequestBudget allowed show two requests, and a read
that drifts to a URL of its own (a mutant whose reader spells
brief_deps=false) fits that ceiling on both handshake shapes. These guards
pin the single fetch instead:

- TestE2E_ShowFetchesTheIssueOnce: the exact requests, path and query, of
  show --json, show --json --id, show --short, show of two ids and text
  show, against both handshake shapes, plus a BD_HTTP_READ_MEMO=0 step
  that must read the row more than once.
- TestDefaultIssueReadsShareOneMemoKey: the resolver's exact probe,
  GetIssue, GetDependenciesWithMetadata, GetLabels and IssueReader().Get
  build one URL, checked without a server.
- TestE2E_CloseRereadsItsOwnWrite: against the MC handshake, closing a
  lone molecule re-reads its own closed row from the server after the
  write, and prints no false "completed molecule" note.
- Corpus shapes can pin `requests:`, counted on the served run alone (not
  setup, verify or {rev:ID} reads). The seven gc and mc show shapes pin
  1, 4, 1, 1, 3, 1 and 3.

Also: TestReadMemoOptInReadsTheRealPollingFlags holds the opt-in to the
real show/list --watch and -w, events tail --follow and bd serve; the
show budget drops from 2 to 1; requestCounter records each request's
query; and ledger row L-read-memo records the memo's bounded difference
from local reads.

Mutation runs against this tree: brief_deps=true on GetIssue and
brief_deps=false on the reader's Get each fail the e2e list and the unit
test; show opting out of the memo fails the e2e list; writes that leave
the memo standing fail the read-your-own-write e2e; a memo keyed by path
alone fails the e2e list; dropping the kill switch fails its unit test
and the e2e step.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(config): list BD_HTTP_READ_MEMO

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* fix(http): memoize a read's answer only once it has decoded

wire.Client.Do stored a memoable read's 2xx before it checked the body,
so an empty or undecodable 200 would have answered every identical read
in the command for the memo's whole TTL with the same failure, instead
of asking the server again. The answer is now stored only after it
decodes; a read that decodes nothing (out == nil) is not stored, and
every memoable read in this package decodes its answer.

TestReadMemoKeepsOnlyAnswersThatDecoded pins it (empty body, truncated
body, undecoded answer). Ledger row L-read-memo says so.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(http): pin the read memo's 5xx, write-method, per-client, in-flight-write and hook-gate behavior

Review follow-ups for the per-command read memo (mc-5dkzp):

- TestReadMemoNeverKeepsA5xx: a 503, 500 or 502 is not memoized and the
  2xx after it is. TestDispatchRetryOfABusyMemoizedReadAsksTheServer
  holds the same through dispatch's retry loop over a real wire.Client:
  the retry reaches the server instead of replaying the 503.
- TestReadMemoIsDroppedByAnyWrite covers POST, PUT and DELETE, not just
  PATCH.
- TestReadMemoIsPerClient: in either order, one client's read is never
  answered from another's memo, and opting one client in opts in no
  other. (A shared memo was otherwise caught only by test-order
  pollution.)
- TestReadMemoDropsAnAnswerReadWhileAWriteWasInFlight pins the clear on
  the write's way back: a read that starts and ends while the write is
  on the wire is not answered again once the write has returned.
- TestHookFiringStoreReReadsOnlyForAnEventWithAHook pins hookConfigured's
  probe: the decorator asks the runner about the event it fires, and
  re-reads the row to build a payload only when a hook for that event
  exists, through a probing double and through a real hooks.Runner.

Each was run against a mutant of the code it pins and fails there:
5xx stored before the status check; only PATCH clearing; one memo
shared by every client; no clear after the write returns;
hookConfigured answering true, false, or about the wrong event.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* docs(changelog): note the per-command http read memo

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-16 04:26:37 +00:00
Julian Knutsen
773b537c75
fix(httpclient): dial a large batchApply under the server's extended budget (#7582)
* fix(httpclient): dial a large batchApply under the server's extended budget

bd serve runs an issues:batchApply over largeApplyItemThreshold (100) items
under DefaultLargeApplyCeiling (5m) on a context detached from the client,
but the wire client capped every request at DefaultTimeout (60s). A large
apply on a slow or loaded server therefore failed client-side with
"Client.Timeout exceeded while awaiting headers" while the server went on
to commit it: an error for an apply that landed.

wire.Request.Timeout now raises (never lowers) a bounded client's
per-request ceiling for one request, through a shallow copy of the
*http.Client that shares its Transport and forced CheckRedirect. ApplyBatch
sets it to LargeApplyTimeout (6m) once the plan exceeds
LargeApplyItemThreshold. The stall-retry deadline uses the same per-request
ceiling.

Tests:
- wire: a slow fake server times out an at-threshold apply under a short
  ceiling and completes an over-threshold one; httpClientFor only raises a
  bounded client and keeps its Transport and CheckRedirect.
- httpapi: TestTheClientLargeApplyBudgetCoversTheServer pins the redeclared
  threshold to the server's and holds LargeApplyTimeout above request read
  + large-slot wait + ceiling + drain grace (5m45s).
- httpclient (served tier): a 101-item apply against a BatchApplier delayed
  DefaultTimeout+5s, dialed through DialWith + DialOptionsForTarget like
  the product, commits and returns success (71s). With the raise disabled
  it fails at 60.03s with Client.Timeout.

Ported from bd-enterprise (gascity/bd-enterprise#195).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(httpclient): pin the raised stall-retry deadline; document the 6m large-apply ceiling

Review fixups for #7582.

- wire: TestARaisedRequestTimeoutAlsoBoundsTheStallRetry pins the one
  OSS-only hunk of the port, the stall-retry deadline read through
  httpClientFor(r). Reverting that line to c.hc.Timeout left the whole wire
  package green; the new case fails on it, because the retry then runs out
  at the client's own ceiling instead of the request's raised one.
- backend/http: Options.HTTPClient said the client is used as given except
  for redirect handling. A batch apply over 100 items now raises a Timeout
  shorter than 6 minutes for that request, so the public doc says so.
- README and CHANGELOG: the HTTP backend's write bound is no longer 60s for
  every request. A batch apply over 100 items gets 6 minutes, and a server
  whose --large-apply-ceiling is raised past about 5m15s can still outlast
  it.
- engdocs/SERVE_RUNBOOK.md: the same limit, next to the stop-grace sizing
  an operator reads before raising --large-apply-ceiling.
- served_batch_apply_test: the race-build comment blamed the client's 60s
  timeout for shrinking the 1000-item apply. A plan that size now gets
  wire.LargeApplyTimeout; it stays shrunk because the embedded tier measured
  it at 903-945s under -race.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-15 23:43:54 +00:00
Bee
b911ed28cd
fix(blocked): climb through a blocked parent with no reason row of its own (#7420)
* fix(blocked): climb through a blocked parent with no reason row of its own

A blocked parent the explained-parent walk meets that neither read returned
was treated as a final unexplained ancestor. That is right for a parent with
a reason row of its own (the complete read decided it), but a parent with
none carries only an inherited bit, and the walk had nothing to decide it
from. In a malformed parent-child cycle (SkipCycleCheck, import, merge
replay) a stale is_blocked = 1 on such a parent held itself up: it dropped
the explained parent below it, which darkened it back, so full repair
converged on the stale state and the doctor check counted 0.

climbBlockedParentsInTx now sorts each unmet blocked parent: explained by its
own subtree (climb on), carrying a reason row and not explained (final), or
carrying no reason row (transparent: entered as explained and climbed
through, and dropBlockedFromAbove decides it from what is above it). The new
read, withoutReasonAmongInTx, uses blockingReasonSQL, the same predicate the
explained-parent reads aggregate.

Closes gastownhall/beads#7386.

Tests: TestParentChildCascade_CycleStaleBitSelfSustains (the issue's repro;
FAIL at 5a5b7cfd66 before the fix, PASS after) and
TestParentChildCascade_InheritedBitStillCascades (a no-reason parent under an
exogenously blocked one still cascades, before and after the fix). Ran the
dolt package's blocked/cascade set and internal/storage/issueops locally
against a dolt 2.2.0 server; golangci-lint 0 issues on both packages.

Agent-Signature: claude-code-claude-fable-5-1-unknown-reasoning on behalf of bee
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* test(blocked): pin the scoped climb through a no-reason parent

Review fixes for #7420 (fixes #7386). All are minors from the adopt-pr
review; no blocker or major was raised, and blocked-state behavior is
unchanged.

- Test evidence: the PR's tests drive only the complete read (recomputeAll
  and the doctor count). The scoped half of the climb, which the per-write
  recompute and merge replay reach through batchExogeneity, had no pin: a
  mutant that skips withoutReasonAmongInTx on a scoped read
  (if !complete { undecided = nil }) passed every blocked test. Add
  TestParentChildCascade_CycleStaleBitScopedRecompute, the same malformed
  cycle driven through RecomputeIsBlockedInTx. It fails at the base and
  under that mutant, and passes here.
- Comment truth in blocked_consistency.go: the set bullet now names every
  parent the walk enters as explained (the read's, those a scoped walk
  finds explained, and the no-reason parents it climbs through) and says
  only the read's are returned; COST counts an edge read per level the
  walk climbs and the reason-row read by id on both paths; the explained
  map's doc covers climbed-through parents; blockingReasonSQL names both
  of its callers.
- Scope of the claim: the CHANGELOG entry and the cycle test's doc now
  say what is fixed, a stale bit that holds itself up THROUGH a parent
  that blocks on its own child, on both the full repair and the per-write
  recompute. A no-reason cycle that holds itself up without running
  through such a parent keeps its stale bits with a doctor count of 0,
  as it did before this change: one above such a parent (Q<->D over P),
  and one with no close gate at all, which an ordinary close can leave
  behind. Those are left to a follow-up, not claimed here.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <julianknutsen@users.noreply.github.com>
2026-10-15 10:41:35 +00:00
Julian Knutsen
9c4b3d8bab
perf(externaldeps): annotate a page's blocking state in one pass, not one IsBlocked per id (#7556)
* perf(externaldeps): annotate a page's blocking state in one pass, not one IsBlocked per id

The external-dependency BlockingAnnotator overlaid IsBlocked on the
backend's answer one item at a time. Each IsBlocked made several reads
(the stored flag, both edge tables, the blockers' statuses), read the
item's own edges again, and resolved its external refs on its own, so a
foreign project was opened once per item. Over a remote serve,
listBlockingAnnotations cost 9-20 ms per id, and a 1,400-id refresh took
about 12 s.

The annotator now reads the whole page once:
- the local half through a new optional storage.BlockersBatchStore,
  IsBlockedWithBlockersBatch, backed by
  issueops.IsBlockedWithBlockersBatchInTx (the stored flags, the blocked
  ids' blocking edges from both tables, one status read). The dolt and
  embedded dolt stores implement it, and a store without it still gets
  one IsBlocked per id;
- the external half from ONE read of the page's own edges, whose refs
  are resolved together, so each foreign project is opened at most once
  per request.

The answer per id is IsBlocked's, unchanged: a blocked item's BlockedBy
is the local live blockers, then the unsatisfied refs. Both reads now
order a source's blocking edges by (depends_on_id, type). The single
read had no ORDER BY, so its order was the table's, which is a UUID key
on these tables.

Tests:
- TestBlockingAnnotatorReadsAPageOnce pins 1 own-edge read, 1 foreign
  open, 1 batch read and 0 per-id reads for 40 ids. On the old body it
  failed with 40, 30, 0 and 40.
- TestBlockingAnnotatorFallsBackPerIDWithoutTheBatchRead covers a store
  without the batch read.
- TestPolicyBlockingAnnotatorBatchEqualsPerItem (embedded dolt) compares
  the batched answer, and IsBlockedWithBlockersBatch, against the
  per-item path by deep equality. It covers local blockers of every
  blocking type, closed and pinned blockers, parent-chain blocking,
  waits-for and conditional-blocks gates on both planes, satisfied,
  unsatisfied and unconfigured external refs, a non-blocking external
  edge, local plus external, a free row and a missing id.
- TestPolicyBlockingAnnotatorForeignOpensAreSublinear logs N=10 at 1.9 s
  per-item vs 0.26 s batched, and N=40 at 7.5 s vs 0.29 s. Foreign opens
  go from N to 1.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(dolt): pin server-mode IsBlockedWithBlockersBatch against IsBlocked

Review apply-fixes for #7556, iteration 1. Two minors are taken here, the
merge conflict is left to the pre-approval rebase, and two items are
declined. All changes are a test and a doc comment; no runtime code moves.

Taken: server-Dolt test for the batch arm (Test Evidence Quality)
- TestIsBlockedWithBlockersBatchMatchesPerRow runs the DoltStore arm that
  bd serve and server-backed bd list take, which no test executed before.
  It compares each id's entry with IsBlocked through reflect.DeepEqual.
  The fixture reuses the embedded fixture's shapes as the raw store sees
  them: blocks, a parent-child chain (the list is nil), waits-for,
  conditional-blocks, a wisp source and a wisp blocker, closed and pinned
  blockers, and external refs (blocks and related), which the raw store
  drops. It also covers a free row, a missing id and repeated ids. Two more
  cases reach code the embedded test does not:
  - one id that exists in both issues and wisps, with an edge in each table;
  - 210 blocked ids with 210 distinct blockers, more than queryBatchSize
    (200).
- Mutants were applied to the PR's code with the test binaries rebuilt.
  Five fail the new test and pass the existing embedded
  TestPolicyBlockingAnnotator* tests:
  - edge-chunk: read only the first edge chunk;
  - status-chunk: read only the first status chunk;
  - no-dedup: drop the repeated-id check;
  - table-order: read wisp_dependencies first;
  - server-wrapper: the DoltStore wrapper returns an empty map.
  A sixth mutant, no-order (drop ORDER BY), fails both the new test and the
  existing embedded batch-equals-per-item test.

Taken: BlockersBatchStore doc (Contract and Interface Fidelity)
- The doc now says the interface answers for the raw store alone, with no
  external-dependency policy applied. It is meant to be called from inside
  that policy's decorator, not probed through UnwrapStore by a caller that
  holds the decorated store.

Deferred to the pre-approval rebase: conflict with #7552 (role_accessors.go)
- This change does not rebase. On a scratch merge with main 70887e708a,
  this resolution was validated:
  - keep this PR's
    state := states[id]; if state.blocked { BlockedBy = state.blockers };
  - nest #7552's IncludeIsBlocked flag (IsBlocked = boolPtr(true)) inside
    that branch;
  - keep boolPtr.
- Results on that merged tree:
  - go vet: clean;
  - go test ./internal/storage/externaldeps/: 66 pass, including
    TestBlockingAnnotatorFlagsUnsatisfiedExternalRef;
  - embedded BlockingAnnotator|IsBlocked tests: 22 pass;
  - served httpclient Blocking|Blocked|ExternalDependency|Ready tests with
    BEADS_HTTP_TEST_REQUIRED=1: 96 pass, plus 4 known-divergence skips that
    predate this change.

Declined
- Duplication (route blockedStates' external half through
  blockingStateFromRecords): this refactors contributor code, adds the
  "external dependencies: resolve blockers:" error prefix, and is not
  test-only. Recorded as a follow-up.
- Pattern Drift nit (move the guards next to their methods, rewrap the
  AnnotateBlocking comment): optional churn next to the conflict region.
  Recorded as a follow-up.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 11:52:37 +00:00
Julian Knutsen
ea0f6e1a46
feat(http): advertise issues.update.claim and route claims by it (S24a/b) (#7555)
* feat(http): advertise issues.update.claim and route claims by it (S24a/b)

updateIssue has served `claim` (upstream #6890), resolved on both planes so
a wisp is claimable, with no capability token: a client could only find it
by eating the skew 400, and gc's wire_compat preflight (which lists
issues.update.claim as an optional token) never saw it, so gc claimed every
wisp through claimIssue, whose Claimer role excludes the wisp plane.

Server: CapIssuesUpdateClaim ("issues.update.claim") joins
behaviorCapabilities. The spec names it on UpdateIssueRequest.claim and in
the context capabilities prose (apigen regenerated, descriptions only). No
request or response member changes; wire_revision stays 2.

Client: Lifecycle.Update reads the token from the cached handshake before
the dial. With it, every claim is one updateIssue call and its answer is
final (no fallback on any error). Without it, a claim alone goes straight to
claimIssue (no doomed 400 first; a wisp still refuses by name,
W-ClaimRequest.Wisp, whose Why now names the token), and a claim beside a
patch, a guard or a force override is refused before the dial with
ErrUnsupported{Capability: issues.update.claim}. serverPredatesUpdateClaim
(the error-sniffing fallback) is gone.

Tests: the four capability lists (routes.go, wire/handshake.go,
server_test, wire vocabulary_test) move together; unit tests for both
routes (pre-dial refusal dials nothing; with the token nothing falls back);
served tests claim a wisp through updateIssue with the token (zero claimIssue
dials, post-claim revision stitched), refuse it by name without it (zero
updateIssue claims, one claimIssue), and refuse a claim beside a patch
without the token while the same request claims and patches the wisp with
it. Live-mutated: ignoring the token fails the unit test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(http): count only ready --claim's requests in the claim-next check

TestE2E_AdvertisedExternalPolicyIsPassedThrough counted every request
since counter.reset(), so its "no by-id claim" check also saw the earlier
`update <other> --claim` step. The MC handshake mask does not advertise
issues.update.claim, so this PR routes that claim to claimIssue
(POST /v0/beads/issues/<id>:claim) by design, and the check failed in
Bazel / embedded Dolt tier (//backend/http:http_served_test shard 6 of
7), turning CI Gate / Required red.

Count from the snapshot length taken just before `ready --claim`, so the
check covers only the requests that command issues. The counter is not
reset: the close-policy check below scans the same history.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(http): refuse a token-less template claim on the client's pre-read

Against a server without issues.update.claim, Lifecycle.Update sends a
claim alone to claimIssue (claimWithoutUpdateClaim) and returns before
applyTemplateGuardForServer runs. claimIssue's role has no template check
(bd-jkp9v3), so `bd update <template> --claim` against every server
without the token claimed the template. The base refused it before any
write: on the client's pre-read against a server without
issues.update.allowTemplate, and in updateIssue's own guard against one
with it.

claimWithoutUpdateClaim now pre-reads the row and returns
AuthorizeTemplateUpdate's refusal before the claimIssue dial, whatever the
handshake says about allowTemplate: that token means updateIssue guards
templates, and this route never dials updateIssue. Any other row goes on
to claimIssue as before (a wisp to its refusal by name), and so does an id
the pre-read cannot find, so claimIssue's not-found stays the one reported.

TestUpdateClaimRefusesATemplateWithoutTheToken (stub wire) and
TestServedUpdateClaimRefusesATemplateWithoutTheToken (served tier) pin it
under an allowTemplate-only and a neither-token handshake. Both go red
without the pre-read; the allowTemplate-only legs also go red if the
existing guard is only moved above the claim branch.

TestUpdateClaimFailsClosedWhenThePreReadFails pins the pre-read's third
arm under the same two handshakes: a read that fails other than as
not-found returns that failure and never dials claimIssue, since the
row's template bit is unknown then. A best-effort pre-read (that failure
falling through to the claim) passed go test ./internal/httpclient/ and
its served Claim|Template tier before this test, and goes red on both
legs with it.

A claim beside AllowTemplate is refused before the dial with
ErrUnsupported naming the token, like a claim beside a patch, a guard or
a force override (isClaimOnlyUpdate already excluded it), but no prose
said so: claimWithoutUpdateClaim's doc, CapIssuesUpdateClaim's doc, the
test doc and the CHANGELOG entry now list it, and a
TestUpdateClaimAgainstAServerWithoutTheToken row pins it. Dropping
!req.AllowTemplate from isClaimOnlyUpdate passed the same unit and served
runs without that row and goes red with it.

The prose now follows the handshake rule too: the W-UpdateRequest.Claim
ledger row (and its re-rendered md) no longer calls claimIssue the
fallback for a server that 400s the member, Update's doc names
claimWithoutUpdateClaim instead of the nonexistent claimRouteForServer,
wire.UpdateFlags.Claim drops the "safe to read as skew" rationale no
caller relies on anymore, and the CHANGELOG entry says a template is still
refused on the token-less route.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 11:47:45 +00:00
Julian Knutsen
70887e708a
feat(http): serve IsBlockedBatch through the BlockingAnnotator role behind dependencies.blocking.isBlocked (#7552)
* feat(http): serve IsBlockedBatch through the BlockingAnnotator role behind dependencies.blocking.isBlocked

gc's native ready cache needs bd's stored, transitive is_blocked column for
the active set (beads.BlockedQuerier.IsBlockedBatch). Over http the client
answered it with the generated stub, so every reconcile logged
`operation "IsBlockedBatch" not supported by the http backend`, the
projection stayed disabled and native Ready took ~5 s on the MC shadow.

No read a deployed (1.3.0) server serves can answer the column exactly:
the /v0 reads hydrate issues without is_blocked, the blocking decoration is
direct and two edge types only, and the flag also propagates down
parent-child edges and reads waits-for / conditional-blocks gates. So this
is an additive server behaviour plus a token, not a client-side
composition.

- issueops: BlockingRequest.IncludeIsBlocked asks for, and
  IssueBlocking.IsBlocked carries, each answered id's stored flag (nil =
  no row, never "free"). issueops.BlockedFlags is the one library entry
  for the batched read over the role, chunked at BlockedFlagsChunk (100,
  held equal to the served issue_id bound by an httpapi test).
- Bodies: storage/issueops.ExecuteBlockingAnnotation reads the flag with
  IsBlockedBatchInTx in the same read transaction (dolt, embeddeddolt and
  every sqlkit-style store that calls it); the unit-of-work body asks the
  new DependencyUseCase/Repository IsBlockedBatch (same shared SQL body).
  externaldeps keeps IsBlockedBatch's external-blocker invariant on the
  local and remote annotators, and its Store.IsBlockedBatch no longer makes
  the raw edge read over a role-served store.
- Wire: GET /v0/beads/dependencies/blocking accepts `is_blocked` behind the
  new behaviour token dependencies.blocking.isBlocked (spec, apigen,
  wire-shape golden). wire_revision stays 2.
- Client: httpBlockingAnnotator forwards IncludeIsBlocked and refuses it
  BEFORE dialing with *ErrUnsupported{Capability:
  dependencies.blocking.isBlocked} against a server that does not
  advertise it; Store.IsBlockedBatch is issueops.BlockedFlags over that
  annotator, so gc's errors.As latch still sees the typed refusal on an
  old server, now naming the token that lifts it.

Tests: conformance case RunBlockingAnnotatorReportsTheTransitiveBlockedFlag
(child of a blocked parent is blocked, free is false, a ghost has no flag,
unasked carries none, BlockedFlags agrees) on the dolt, embedded, uow and
served legs; httpapi handler and chunk-bound tests; client tests for
chunking, the wire shape and the zero-dial refusal; vocabulary/capability
gates.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(externaldeps): keep the external blocker on the unit-of-work IsBlockedBatch

bd serve on a server or proxied-server workspace serves through the
unit-of-work arm (WrapUOWProvider) and advertises
policy.external_dependencies, so a client trusts the server's answers and
adds no external term of its own. That arm's dependency use case did not
override IsBlockedBatch, so the new dependencies.blocking.isBlocked flag
answered the stored is_blocked column, which has no external term. An issue
blocked only by an unsatisfied external: ref came back is_blocked=false,
with that ref listed in its own blocked_by, while GET /v0/beads/ready
excluded it.

dependencyUseCase.IsBlockedBatch now raises every id that has an
unsatisfied external blocker, deciding by the edges and resolver
GetReadyWork excludes by, read through the same unit of work. Unlike Ready
it resolves only the refs of the ids the column answered: a client asks in
chunks, and no chunk should reopen every foreign project the workspace
names. An id naming no row stays absent.

Also correct the store decorator's IsBlockedBatch doc (the served flag now
comes from the annotator role), add the CHANGELOG entry for the new
capability, and pin the flag against Ready on every arm: WrapUOWProvider,
directly and through its BlockingAnnotator (an absent id stays absent), the
local annotator, and the served test, which now asserts the flag and the
batched read. A counting opener pins that the unit-of-work read opens only
the foreign projects its asked rows name.

Checked against a real bd serve on a proxied-server workspace. At
97bcb57738 an issue blocked by external:remote:payments answered
is_blocked=false while /ready excluded it. With this change it answers
true, and once the capability ships /ready includes the issue and
is_blocked is false. Asked about a free issue alone, the server resolves
no external project; asked about the remote-blocked issue, it resolves
remote and leaves alone the unconfigured project another issue names.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(externaldeps): pin that a satisfied external blocker leaves the unit-of-work is_blocked false

Review of 192b73676 (bee-ghosttrack): the two new IsBlockedBatch tests only
had unsatisfied external refs, so a body that raised every answered id with
any external record (ignoring the resolver) stayed green.

TestWrapUOWProviderBlockedFlagResolvesOnlyTheAskedRefs's opener now
returns a foreign store whose labelled index carries a CLOSED provider of
provides:payments (the decorator_test.go shape). be-shipped blocks on
external:remote:payments (satisfied) and must read false; be-blocked moves
to external:remote:invoices (still unsatisfied) and reads true; both still
open /projects/remote once. With the reviewer's mutant applied
(`if _, ok := answered[id]; ok && len(state.refsByIssue) >= 0 {`) the test
fails on be-shipped in both new cases; restored, it passes
(bazel test //internal/storage/externaldeps:externaldeps_test).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(externaldeps): pin the store arm's satisfied external ref and the unasked is_blocked flag

The store arm decides an external blocker in two places: IsBlockedBatch's
local loop skips a ref the resolver reports satisfied, and the
BlockingAnnotator role folds IsBlocked's verdict, which applies the same
filter, into the served flag. Both annotators, store and remote, add that
flag only when the request asks for it. No test checked either store answer
for a ref whose project had shipped the capability, or that an unasked
annotation carries no flag, so dropping any of those checks left the
package green.

Pin them: IsBlockedBatch and the store annotator answer false for a row
whose only blocker is external:remote:payments once remote has closed its
provides:payments issue, and neither annotator flags a row unasked, even
one an external ref blocks. This is the store half of bd-yewju3; the
previous commit pinned the unit-of-work half.

Each mutant below survives the package at the previous commit and fails
exactly one test here (go test -tags=gms_pure_go
./internal/storage/externaldeps/; unmutated, bazel test
//internal/storage/externaldeps:externaldeps_test passes):
- decorator.go IsBlockedBatch: `&& !satisfied[dep.DependsOnID]` made
  `&& len(satisfied) >= 0` fails TestIsBlockedBatchHonorsShippedExternalRef.
- decorator.go IsBlocked: `if !satisfied[ref.raw] {` made always true fails
  TestBlockingAnnotatorLeavesShippedExternalRefUnblocked.
- role_accessors.go: `if req.IncludeIsBlocked {` made `if true {` fails
  TestBlockingAnnotatorAddsNoFlagUnasked.
- remote_roles.go: dropping `req.IncludeIsBlocked &&` fails
  TestRemoteBlockingAnnotatorMergesExternalBlockers.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 08:46:13 +00:00
Julian Knutsen
eaed71fbba
fix(http): bd promote over http through issueops.Promote and issues.update.persistence (S21b) (#7554)
* fix(http): bd promote over http through issueops.Promote and issues.update.persistence (S21b)

`bd promote` refused over a connected HTTP workspace: it called the raw
PromoteFromEphemeral, which the http backend stubs, and updateIssue's patch
did not publish persistence. MC's PR-intake formula
(mol-normalized-issue-pr-v1) runs `gc bd promote <id> --reason ...`, so this
blocked the HTTP cutover.

- One library entry, issueops.Promote(ctx, PromoteRoles{Reader, Lifecycle,
  Commenter}, PromoteRequest{Actor, IssueID, Reason, Provenance}): read the
  row, refuse a non-wisp (*NotAWispError, the sentence bd promote always
  printed), move it with ONE Lifecycle.Update whose patch is Persistence =
  persistent (the role's atomic plane move: same id, labels, deps, events and
  comments carried), then write the promotion comment (a comment failure is
  reported, the promotion stands). A no-op move is a lost race and reports
  NotAWispError without a second comment.
- cmd/bd: the direct route runs it over the store's own role accessors, so a
  connected http workspace and a local one take the same entry. The history
  label is "bd: promote <id>" where the store carries caller labels.
- bd serve: IssuePatchBody.persistence (persistent|ephemeral|no_history,
  anything else a 400 naming patch.persistence) on PATCH
  /v0/beads/issues/{id}, behind the new additive behavior token
  issues.update.persistence. wire_revision unchanged. Spec, apigen and the
  wire-shape golden regenerated; the handler only decodes and passes it to
  the role.
- httpclient: Lifecycle.Update sends patch.persistence only to a server
  advertising the token, and otherwise refuses BEFORE the dial with
  *wire.CapabilityError naming it (unwraps to *storage.ErrUnsupported with
  Capability set). The batchApply item still refuses it
  (W-IssuePatch.Persistence now reads "ON issues:batchApply").
- Conformance: RunLifecycleUpdatePersistentPromotesAWisp on dolt, embedded,
  uow and the served http leg; the served
  PersistentPreservesUnversionedClass case is unparked.
- Served e2e TestE2E_PromoteOverHTTP: the mc-promote shape end to end
  (label kept, comment written, one PATCH, re-promote refused), and under
  the MC server mask a pre-dial refusal naming the token with no write.
- Corpus: mc-promote stays pending, relabelled SRV-1 with the reference
  profile's exact refusal (the 1.3.0 server lacks issues.update.persistence).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(promote): run the proxied-server route through issueops.Promote (bd-4y73m9)

The adoption review found (major) that the proxied-server route still
promoted through the old engine, IssueUseCase.PromoteWisp in one
transaction with the comment, while the PR's docs said every route runs
issueops.Promote. Converge rather than reword: runPromoteProxiedServer
resolves the id, then runs issueops.Promote over the provider's own
notifying role accessors (proxiedIssueReader, proxiedIssueLifecycle,
proxiedCommenter), the entry the direct route runs. Both routes now share
promoteFailure (the error mapping) and reportPromotion (the comment
warning and the output), so neither contract can drift.

What a script sees change on the proxied-server route:
- the move is the persistence branch of Lifecycle.Update, so the child
  counter moves with the row: a promoted parent's next
  `bd create --parent` mints .3 after .1/.2 instead of failing with
  "issue already exists";
- on_update fires twice, for the move and for the comment, as on the
  direct route (it fired once, for the comment);
- a promotion is two Dolt commits, "bd: promote <id>" and
  "bd: comment <id>" (it was one);
- a comment that cannot be written leaves the promotion standing behind
  a warning, as on every other route, instead of rolling it back;
- a dual-plane id is refused rather than merged, the update role's
  template guard applies, and a lost race reads "<id> is not a wisp
  (already persistent)".

The PromoteWisp exemption in TestRecordingIssueUseCaseCoversItsSurface
stays, with its reason corrected: no route calls PromoteWisp now, and
its DoltStorage twin fires no hook either. The doc comments on the old
engine (PromoteFromEphemeral on DoltStore and the db layer, PromoteWisp)
no longer say bd promote uses it.

CHANGELOG: the promote entry now names what a script can see on the
direct and proxied-server routes: the hook firings, the comment row
`bd comments` lists, the child-counter carry, the two-commit shape, and
a comment failure that warns instead of rolling back.

Tests:
- TestEmbeddedPromoteMovesAWispParent promotes a wisp parent with .1/.2
  children, a dependency and a comment under a counting on_update hook:
  the hook fires twice, the dependency and the children's parent edge
  survive, `bd comments` lists the old comment and the promotion reason,
  and the next child is .3. At the pre-PR base it fails (the hook fired
  0 times).
- The proxied journey pins the two commits, and new subtests pin the
  child counter (the child_counters row, next child .3) and the two hook
  firings. Against the contributor head each fails: one commit, no
  counter row and a failing create, one firing.

Review iteration 2:
- TestEmbeddedPromoteMovesAWispParent gets its row in the embedded cmd
  shard manifest (shard 64 of 100, from the generator's incremental
  --write). Without it //scripts:go_test_sources_test fails both the
  NotStale and the bare-write guard, and the generator's --check fails.
- notifying.go no longer lists the promotion comment as a divergence.
  The direct route now writes that comment through st.Commenter(),
  which HookFiringStore wraps, so an embedded promote fires on_update
  for it (TestEmbeddedPromoteMovesAWispParent counts both firings). The
  list counts four, and the empty-commit note becomes the fifth item.
- batchapplier.go no longer says persistence is absent from both patch
  bodies: the update body publishes it behind issues.update.persistence,
  and only batchApply refuses it (W-IssuePatch.Persistence).
- Three comments still tied promotion to the old engine: the
  conformance fixture's PromoteWisp no longer claims to do what
  `bd promote` does, the CLI differential's promote scenario names
  issueops.Promote, and export_source's wisp-relations note names
  issueops.MoveIssuePersistenceInTx.
- The Promote unit test compares the whole update patch with
  reflect.DeepEqual. The helper it replaces checked 9 of the patch's 22
  members, so a Promote that also set Description passed it.

Review iteration 3:
- A refused persistence move answered 500 `internal` over http and
  emitted request_error, though the caller could fix the request. The
  role now returns two refusals as ErrValidation, so bd serve answers
  400 invalid_argument and the client maps that back to ErrValidation:
  an unversioned row asked for ephemeral or no_history (the role wraps
  types.ErrUnversionedWispMove, a new sentinel NormalizePersistenceMode
  returns for that case alone, with its message unchanged), and a
  demotion of a row with retained snapshots. A row that breaks its own
  persistence invariants still answers 500.
- RunLifecycleUpdateWispModesRefuseAnUnversionedRow pins the refusal on
  every leg (dolt, embedded dolt, the unit of work, served http): both
  wisp modes match ErrValidation and leave the row durable, unversioned
  and off the wisp plane. Without the wrap the dolt leg fails, and the
  served leg fails on "answered 500 internal".
- TestMoveIssuePersistenceInTxRefusesDemotionWithRetainedSnapshots is
  the first test of the snapshot refusal, and fails without its wrap.
  The unversioned unit test now requires both sentinels. The types tests
  pin the sentinel, its unchanged message, and that no row-invariant
  error matches it: checking the unversioned class ahead of the
  invariant checks fails them.
- CHANGELOG: the promote entry counts five script-visible changes, the
  fifth being the lost-race refusal sentence, and says the two refusals
  now answer 400 over bd serve and gain a `validation failed: ` prefix
  locally.
- NotAWispError's doc comment records that a no-history wisp is refused
  too: the gate reads the Ephemeral flag, as bd promote always has, and
  `bd update --persistent` is the move for that row.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 05:51:53 +00:00
Julian Knutsen
048f5e5192
fix(httpclient): bound response-header waits and retry stalled reads on another connection (#7553)
* fix(httpclient): bound response-header waits and retry stalled reads on a new connection

The maintainer-city shadow run saw 22 requests sit out the full 60s client
timeout, each on one silent HTTP/1.1 connection while other connections to
the same serve kept answering.

wire.Client now bounds an idempotent request's wait for response headers
(15s default; BEADS_HTTP_RESPONSE_HEADER_TIMEOUT or
DialOptions.ResponseHeaderTimeout; "off" disables), separately from the 60s
overall timeout, and stops the timer once headers arrive so a streaming body
is never cut. A read (GET/HEAD/OPTIONS, or a Request marked Idempotent: the
POST batch get) that stalls or has its connection reset is retried once: the
stalled attempt's cancellation closes its connection and idle connections are
dropped, so the retry dials fresh. Every retry is counted
(wire.StallRetryCount) and reported (OnStallRetry; the dialer's default logs
one stderr line with the count).

Writes are never retried and get no header timer, because bd serve answers
a write only after it commits (a large batch apply can legitimately take
tens of seconds). A write whose overall timeout fires before its headers
fails with a typed *wire.ResponseHeaderTimeoutError (inside *ConnectError)
naming the host and saying the outcome is unknown.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(httpclient): retry a stalled read within the overall timeout (bd-p1wzpi)

Adoption review of #7553 found three defects in the header timer and its
retry. This maintainer fixup sits on top of the contributor commit and
supersedes three of its claims: that the timer bounds a read's every
attempt, that idle connections are dropped before the retry, and that any
write whose overall timeout fires before its headers is typed.

A slow but healthy read is no longer cut short. The 15s timer cannot tell
a silent connection from a serve still working on the read (serve writes a
read's headers only after waiting up to 10s for a database slot and running
the query), and the retry ran under a fresh 15s timer, so a read slower
than 15s failed at about 30s after sending two requests where it used to
succeed within 60s. The timer now bounds only the first attempt. The retry
has no header timer and runs under a deadline of when the first attempt was
sent plus the client's overall Timeout, the moment that Timeout itself
counts from, so it finishes such a read and no read waits longer than the
60s it could before. The request is built and stamped once and each
attempt sends a copy of it, so the credential source is consulted once,
before any clock starts, and a cold token command costs the retry none of
its budget. What is left: a read whose headers take longer than the timer,
and whose retry then needs more than what remains of the 60s (about 45s)
for its headers and body together, now fails at 60s where it used to
succeed, and the server also gets the abandoned first request. A read whose
first attempt already spent the whole Timeout (the timer off) is not
retried at all and keeps its typed timeout.

A request is typed *ResponseHeaderTimeoutError only once it was written.
Any Timeout() error on a write used to be typed, including the baseline
transport's dial and TLS handshake timeouts where nothing reached the
server, and the message then said the write "may or may not have been
applied", gave the overall timeout as the wait, dropped the transport's
error, and Temporary() was true. The timer is now armed from httptrace's
WroteRequest (where http.Transport.ResponseHeaderTimeout starts its own
clock), which also records that the request went out. A timeout is typed
only for a written request whose caller's context is still alive; a dial
or TLS timeout stays a plain *ConnectError. After is the measured wait,
Err is printed, Temporary() reports Replayable (false for a write), and a
retried read says why its first attempt was abandoned (Reason). Arming at
the write also stops a slow credential source (BEADS_HTTP_TOKEN_COMMAND may
take 30s) from being reported, counted and retried as a server stall.

No idle-connection flush. CloseIdleConnections ran on a transport shared
by every no-CA dial in the process (baselineTransport), so one stalled read
dropped idle keep-alives to every host. Canceling the stalled attempt
already closes its connection (net/http never pools a canceled HTTP/1.1
connection), so the retry rides another one.

Also:
- wire.Request.Replayable (GET/HEAD/OPTIONS, or Idempotent) is now the one
  predicate behind both the wire's stall retry and dispatch's bounded 503
  retry, so the batch get, a read on POST, is retried on a 503 too.
  dispatch's doc comment no longer says the wire never resends.
- The docs, the error and the stall line say the retry goes out on
  "another connection", not a "new" one: the idle pool may hand it one
  already open. dialStream's doc says the events:watch connect gets
  neither the timer nor the retry.
- BEADS_HTTP_RESPONSE_HEADER_TIMEOUT takes bare seconds ("30") as
  BEADS_GATE_WAIT_TIMEOUT does, and "off" in any case. "off" or 0 turns off
  only the timer: a reset connection is still retried and a written request
  that times out is still typed.
- README says "each read" and spells out the above, bounding the retry by
  its headers and body together; it no longer says a streaming body is
  never cut off (the header timer never does, the 60s bound still does).
  The CHANGELOG entry moves from Added to Fixed and states the same bound.
- CredentialProvider's doc, in wire and on the public bdhttp alias, says
  Authorize runs once per request, not once per attempt: the stall retry
  resends the stamped headers up to the overall timeout later, so a
  single-use or short-lived credential may draw a 401 there, which reaches
  Refresh like any other.

Tests: a read queued behind a LimitListener backlog finishes once a slot
frees; a read slower than the timer on every connection finishes on its
retry; a slow credential source is not a stall, and one that takes the
whole Timeout to resolve still leaves the retry its budget (starting the
deadline before the stamp, or re-stamping the retry, fails it); a write
whose dial times out is a plain *ConnectError; a stall leaves another
server's idle connection on a shared transport alone; a retry after a
reset names the reset; the stall-twice case is bounded by the overall
timeout and pins After, Reason and the message; dispatch retries a read
on POST on a 503, and GetMany's batch get is sent as one (dropping its
Idempotent fails that); the env knob takes "30", " 1.5 " and "OFF" and
refuses "-3". A read whose first attempt spent the whole Timeout is not
retried and keeps its typed error (dropping the spent-budget check fails
it; no other test did). A stall retry that draws a 401 refreshes once and
authorizes only the re-issued request, so Authorize runs twice, not three
times (re-stamping the retry fails it).

Not covered, as before, and now said in README and the CHANGELOG: the
events:watch stream's connect (dialStream) has no header timer or retry.
Nor is a reset on Windows: stallReason matches syscall.ECONNRESET, which a
Windows reset (WSAECONNRESET) is not, so there a reset connection is
retried only when it surfaces as EOF.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Eddie the Engineer <adopt-pr@gascity.com>
2026-10-14 05:04:19 +00:00
Julian Knutsen
e7ad85bcff
fix(http): bd export over http through issueops.ReadExport (S21) (#7551)
* fix(http): bd export over http through issueops.ReadExport (S21)

bd export read through raw SearchIssues and the bulk label, dependency and
comment reads, which the HTTP backend refuses (B3 #16), so gc's JSONL export
order (gc bd export --all) failed on every remote scope.

issueops.ReadExport composes role reads every deployed server publishes: an
unlimited Reader.List for rows, labels and counts, EdgeReader.ReadEdges for
dependency records, and Reader.Get with comments for rows that have any. The
remote route feeds it into the same runExportFromSource the direct and
proxied routes use. A failed read fails the export. gc-export-all passes
against the reference profile; mc-promote stays pending (S21b) with its exact
reason: it needs a new server capability.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(export): read remote memories from the memory plane, pin unmarked no-history rows (bd-o4qlui)

Review fixes for #7551, attempts 1 to 3.

Attempt 3. No test reached roleExportSource.ExportMemories's error
path. A unit test now drives it with a fake store: a memory role that
cannot be reached, and a listMemories that fails, each end the export
with that error and no map. Rewriting either branch to degrade to an
empty map would bring back a remote `bd export --all` that exits 0 with
no memory in it.

Declined in attempt 3, for finalize to file: an unlimited HTTP listing
that stops on a has_more page with no next_cursor returns the rows it
has with HasMore false, so ReadExport's refusal cannot fire over HTTP.
Only a server that breaks the spec's has_more/next_cursor pairing gets
there; this repo's server always sends a cursor with a non-empty
has_more page. The walker (fetchIssuePages) is unchanged since #7247
and shared with bd list, so the fix needs its own PR that checks bd
list's callers first. The remote route also has no single snapshot: a
comment changed between List and Get can make the written count
disagree with the bodies kept. That belongs in one follow-up with the
bounded retry declined in attempt 1, as a count/body consistency check.
The --include-infra e2e leg stays declined. The wire does carry the
knob: the client sends include_infra and include_templates on
listIssues, and the server decodes both. The GetInfraTypes nil-store
guard stays declined, and auto-export's own GetAllConfig memory read
predates this PR.

Attempt 2. A remote export dropped every memory. Over HTTP, `bd export
--all` and `--include-memories` exited 0 without a memory record, where
the base refused with exit 1. The memory lines came from GetAllConfig,
which over HTTP is listSettings, and the server's settings enumeration
withholds every kv.* row, memories included. gc's JSONL export runs
`gc bd export --all`, so its archive lost them without a word.

The exportSource seam now asks each source for its memories
(ExportMemories), not for the whole config table. The local and proxied
sources still take them from their config table. The remote source reads
the served memory plane, store.Memories().List (listMemories), which the
deployed bd 1.3.0 server advertises as memories.list. httpMemories.List
refuses a page that reports has_more: ListResult has no cursor, so a
truncated plane would pass for the whole one. The memory error now says
"failed to read memories"; nothing in beads or gascity matches the old
text.

TestE2E_ExportOverHTTP seeds two memories with `bd remember` over HTTP.
It requires exactly those two in --all, in --include-memories and in the
local re-export of the imported file, and none in the default export. A
unit case pins the has_more refusal. The CHANGELOG entry says where
remote memories come from, so its "Nothing is dropped" holds. The
WispPlaneIDs doc no longer says the listing refuses an id set, which
stops being true once a server advertises issues.list.ids (#7550).

Attempt 1. The remote route's WispPlaneIDs doc and the CHANGELOG had the
effect of the missing wisp_plane marker backwards. They said import then
routes a no-history row by its flags. A current import does the
opposite: it pins an unmarked no-history row to the durable plane
(applyImportWispPlane). So a genuine no-history wisp in a remote export,
which gc's session, wait and order beads are by default, re-imports onto
the Dolt-versioned issues plane, and gc's JSONL archive keeps it.

The route still stamps nothing. No served read publishes plane
membership: listIssues rows carry no plane field, however the listing is
scoped. Inferring the marker from the flags would re-plane a durable row
that still carries the flag and drop its relations (bd-r9uce), and
refusing the export would re-break gc's remote JSONL export. The doc and
the CHANGELOG now state the real consequence, and serving the membership
is bd-1lbvas. TestE2E_ExportOverHTTP pins it. A no-history row is
exported without the marker by --all and by the default export. A fresh
local workspace imports it onto the durable plane, against a locally
created control row that the local route does mark.

Attempt 1 minors and nits: ReadExport refuses a listing page marked
HasMore, as the uow source does, and unit cases pin that and a failed
comment read. The exportSource interface doc names ExportIssues and its
contract. The export error now says "failed to read the export set"
rather than "failed to search issues"; nothing in beads or gascity
matches the old text. roleExportSource.GetInfraTypes says it is unused
on this route.

Declined, for finalize to file: a bounded retry of ReadExport when a row
vanishes between the reads (a design change; today's fatal behavior is
pinned), and e2e rows for pinned, gate, template and infra-typed issues,
including an --include-infra leg over HTTP. Also declined: a nil-store
guard on roleExportSource.GetInfraTypes. The method is dead on this
route, and the source's other store readers (ExportIssues, GetConfig,
ExportMemories) dereference the store unguarded too.

Validation, attempt 3: git diff --check, the go.mod-pinned gofmt, go
vet, gazelle (no change) and the go_srcs check. The new test passed with
no skips. Three mutants were killed. An empty map with no error on a
failed listing, and the same on an unreachable role, each fail their own
subtest. An empty map returned beside the listing error fails the no-map
assertion; export.go discards the map on error today, so that case pins
the interface's "fail rather than return a partial set" contract, not
visible output.

Validation, attempt 2: git diff --check, gofmt and go vet. The
httpclient memory tests passed, 23 of them: 4 units and 19 served. So
did TestE2E_ExportOverHTTP on embedded Dolt with
BEADS_HTTP_TEST_REQUIRED=1 and no skips. Two mutants were killed: a
remote source that reads the settings listing fails 3 e2e assertions,
and dropping the has_more refusal fails the unit case. Export parity
passed with no skips on the classic (memory determinism, byte
stability), embedded and proxied tiers, including the proxied-versus-
classic byte identity with memories. The PR is now behind main, which
added #7549 and #7550, and conflicts with it textually in CHANGELOG.md,
backend/http/BUILD.bazel and issueops/BUILD.bazel. With those resolved
by union, the merge with main e6a891b900 builds and vets. On that merge,
in CI's pinned-shard layout, the served http tier passed 49 of 49,
including #7550's list-scope e2e. The command corpus ran 128 shapes: 113
pass, 3 allowed, 12 pending and 0 fail. gc-export-all passes without its
pending mark, the one shape that moved. The issueops ReadExport and
ListGates tests passed, as did the embedded and proxied export tiers.

Validation, attempt 1: the issueops ReadExport tests, with two mutants
killed; TestE2E_ExportOverHTTP with a flag-stamping WispPlaneIDs mutant
killed on three assertions; the cmd/bd export tests in a scrubbed env.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 03:59:35 +00:00
Julian Knutsen
e6a891b900
feat(list): bd list --no-assignee, --title and --id over http behind per-parameter tokens (S15b) (#7550)
* feat(list): bd list --no-assignee, --title and --id over http behind per-parameter tokens (S15b)

Over http `bd list --no-assignee`, `--title` and `--id` refused: the list
encoder had no listIssues parameter for ListRequest.NoAssignee,
TitleSearch or IDFilter (B3 #11). All three are maintainer-city cutover
blockers: gc's default on_boot hook runs
`bd list --no-assignee --status=in_progress ...` on every agent boot, the
PR-intake formula lists by `--title "$SUBJECT"`, and the pr-review skill
by `--id <ids>`.

The library entry is unchanged: issueops.Reader.List, whose bodies
already apply all three through workapi.BuildListFilter, the same entry
the local route reads through.
- Server: GET /v0/beads/issues decodes `no_assignee`, `title` and `id`
  into those ListRequest members (`id` handed over as written; the role
  splits, trims and de-duplicates, as countIssues already does). Each is
  behind its own new behavior token, issues.list.noAssignee,
  issues.list.title and issues.list.id (additive, wire_revision stays 2):
  each is a separate caller's dependency, and a refusal names exactly the
  one it needed. The spec parameters cite their tokens; apigen and the
  wire-shape golden are regenerated.
- Client: the encoder maps the three members to their parameters, and
  httpReader.List refuses a request carrying one BEFORE dialing, with
  ErrUnsupported{Capability: <token>}, unless the handshake advertises
  that token (refuseUnservedSearch generalised to one table,
  tokenedListFilters, with bd search's `q`). Never a round trip an older
  server answers unknown_parameter, never a wider listing.
- ErrHTTPUnsupported's text names the capability when it has one, as the
  portable sentinel does, so `bd list` says which token the server lacks.
- The three E-ListRequest ledger rows retire; the ledger is re-rendered.

The id scope unparks the served Reader contract: the sixteen cases parked
on E-ListRequest.IDFilter now run against the in-process server. Twelve
pass as written; four still park, each now on the member it actually
needs (Offset twice, PinnedFlag, ExcludeTypes). The served refusal sweep
asserts SpecPrefix where it asserted IDFilter.

Tests (failing first): handler test that each parameter reaches the
storage filter; client tests that each member rides its parameter behind
its token and refuses pre-dial, zero dials, naming its token, against a
server lacking it; round-trip and vocabulary gates; capability lists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(backend/http): the four maintainer-city list shapes end to end; corpus marks relabelled SRV-1 (S15b)

TestE2E_ListScopeFiltersOverHTTP runs the corpus shapes gc-list-on-boot,
gc-list-on-boot-workflow, mc-list-title and mc-list-id through the real
bd against the in-process server (which advertises the three new tokens)
and requires each to answer an embedded workspace's rows and exactly the
rows it names: --no-assignee leaves out an assigned in-progress row on
the same lane, --title is case-insensitive, --id lists a closed id under
--all. Against the MC production server's handshake shape each shape
exits non-zero naming its token and never dials listIssues.

The corpus suite masks the served side to reference_server.json, the
deployed bd 1.3.0 surface, which advertises none of the tokens, so the
four shapes now get the named pre-dial refusal there. Their marks move
from S6c+S15b to SRV-1 with that reason, like mc-search. With a scratch
profile that adds the three tokens and parameters (not committed) exactly
these four shapes report "now passes; remove the pending mark" and
nothing else changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(list): adopt bd-enterprise's listIssues `id` contract: repeatable, issues.list.ids (S15b)

bd-enterprise already publishes listIssues `id` (E9-S2): a REPEATABLE
parameter, one id per value, behind `issues.list.ids`, with its client
sending IDFilter split into one value per member. The first cut of S15b
took ONE comma-separated `id` behind `issues.list.id`, so the same
parameter name had two incompatible meanings: an OSS client's "a,b" is a
400 on an enterprise server, and the enterprise rebase onto main would
collide. OSS now publishes enterprise's contract unchanged:
- Server: listIDFilter (ported verbatim) decodes the repeated `id` into
  ListRequest.IDFilter's comma set; more than 200 values, an empty value,
  a comma, leading or trailing whitespace, a control character or an
  over-long id is a 400 `invalid_argument` on `id`. The spec parameter is
  enterprise's (array, maxItems 200); apigen and the wireshape golden are
  regenerated.
- Token: CapListIDs = "issues.list.ids", the enterprise spelling and name
  on both sides.
- Client: the encoder splits the set the way the role does (splitIDSet)
  and sends one `id` per member; the pre-dial gate fires only when the set
  names an id.
- Tests: list_ids_test.go is enterprise's id cases (the exact set reaches
  storage, the 200 bound is served, every refusal beside its control, the
  spec bound matches the constant); the handler, client, round-trip and
  e2e cases move to the repeated shape and the new token.

`no_assignee` and `title` are unchanged. Served tier, http e2e (including
TestE2E_ListScopeFiltersOverHTTP) and corpus_served_test pass; with a
scratch reference profile advertising the three tokens exactly the four
S15b shapes report "now passes".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 00:18:00 +00:00
Julian Knutsen
ec55baa425
fix(http): bd show --current, recompute-blocked and search over http through issueops entries (S6e/S25/S26) (#7528)
* fix(show): --current through issueops.FindCurrentIssue over the Querier role; no swallowed refusal (S6e)

Over http `bd show --current` printed the WRONG issue, exit 0.
resolveCurrentIssueID asked the raw SearchIssues method for the actor's
in_progress and then hooked work; the http backend refuses that method, the
refusal was taken for "nothing found", and the command fell back to the
last-touched issue. The proxied route had its own copy of the lookup, read
inside its unit of work, with the same swallow.

One library entry now: issueops.FindCurrentIssue(ctx, Querier, actor). It
asks the Querier role (queryIssues over http, which every deployed server
publishes) for `status=<s> AND assignee="<actor>"`, limit 1 — exactly the
raw filter, answered in the store's default order, so the same row comes
first on every backend — in_progress, then hooked. A failed read is an
error; "" (two reads that answered and found nothing) is the only miss. The
actor is a quoted, escaped query value; the query language reads
`assignee=none`/`null` as "unassigned", so such an actor (or an empty one)
is not asked about.

Both CLI routes call it with their route's Querier (store.Querier() or the
proxied provider's) through one cmd/bd helper; what stays in the CLI is the
actor and the last-touched fallback (a workspace file). The proxied route
resolves --current before it opens its unit of work, and its duplicate
in-transaction lookup is deleted.

Tests: issueops unit tests (order, failure is an error on either read,
quoting, unaskable actors, each expression evaluates to exactly the raw
filter); cmd/bd fallback tests; served e2e TestE2E_ShowCurrentOverHTTP
(in-progress beats a later-touched issue, hooked when nothing is in
progress, last-touched only when neither). Against the old code the e2e
fails with exactly the reported bug.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(recompute-blocked): one library entry, storage.RecomputeBlocked; a no-op on a remote backend (S25)

`bd recompute-blocked` refused over http ("storage backend does not support
is_blocked recompute"), and gc's BdStore.RecomputeBlocked issues it.

is_blocked is server-derived. Every write a remote client makes is committed
by the server's own store, which keeps the flag under
issueops.BlockedStateInvariant's local-write clause (targeted, inside the
write's transaction; a Postgres-family server has no full recompute at all
and runs a no-op readiness strategy for the same reason). A client cannot
leave the flag stale and holds no copy of it. The invariant's merge clause
(a recompute skipped after a pull, bd-6dnrw.37; a migration that over-set
it, beads#7037) happens to the SERVER's database, and its repair is this
command in the server's workspace.

So the decision is in the library, not the CLI: storage.RecomputeBlocked
(internal/storage/blocked_recomputer.go) is the one entry every backend goes
through. A local store (storage.BlockedRecomputer) recomputes its column; a
store implementing storage.ServerMaintainedBlockedState — the http client,
whose BlockedStateMaintainedByServer answers true — recomputes nothing and
reports MaintainedBy "server". No wire operation: a whole-column rewrite any
bearer of a workspace credential could trigger, for state no client can have
caused, is not one to publish. The proxied route keeps its pinned-connection
plumbing around the same versioncontrolops body the stores run.

--json keeps rows_corrected (gc requires the member; wh-bridge-sync reads the
exit code) as 0 and adds maintained_by: "server"; the text says where the
repair runs.

Tests: storage unit tests (local repair runs, server-maintained wins even
over a local repair and runs none, neither refuses); served e2e
TestE2E_RecomputeBlockedOverHTTP (exit 0, JSON shape, text, ready set
unchanged). With the server-maintained branch disabled the e2e fails with
the old refusal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(search): every bd search route through Reader.List with issueops.SearchListRequest; listIssues `q` behind issues.list.search (S26)

`bd search` called the raw SearchIssues method with its text, which the http
backend refuses ("partial-id search is not available over HTTP"), and its
direct and proxied routes each mapped the flags onto a storage filter by
hand. No served operation could carry search's meaning (title OR id contains
the text, case-insensitively; an id-shaped text also matches exact id, id
prefix and external ref): TitleSearch is title-only and the query language
has no title-or-id predicate.

The library entry is issueops.Reader.List:
- issueops.ListRequest gains Query, SearchIssues' free-text argument as a
  member. Both Reader bodies (workapi/storereader and the uow reader) pass it
  to SearchIssuesWithCounts where they passed "". It narrows and lifts no
  default exclusion; the ready arm refuses it (readyScopeFields).
- issueops.SearchListRequest(query) is search's scope as a request: Query,
  every status (AllFlag), every kind of row (the four Include*), and the
  store's default order (SortBy priority), so --limit keeps the rows a raw
  read kept.
- CLI: ONE runSearch for the direct, proxied and remote routes. One builder
  (searchListRequest) maps the flags onto SearchListRequest; the route only
  picks the reader (store.IssueReader() or the proxied provider's).
  --sort/--reverse reorder the page as before. search_proxied_server.go and
  its hand-built filter are deleted, and search leaves the forbidigo
  filter-writing exemption.
- Server: GET /v0/beads/issues decodes `q` into ListRequest.Query and calls
  the same Reader role, behind the new issues.list.search behavior token
  (spec parameter cites it; additive, no wire_revision bump; apigen and the
  wire-shape digest regenerated).
- Client: the encoder maps Query to `q`; httpReader.List refuses a request
  carrying Query BEFORE dialing, with ErrUnsupported{Capability:
  issues.list.search}, unless the handshake advertises it — never a round
  trip an old server would 400, never the unsearched listing passed off as
  the search. CapListSearch mirrors the server token. Flags listIssues
  publishes nothing for are put on the request so the encoder refuses them
  rather than widening the search.

Tests: Reader contract case RunReaderListQueryMatchesTheTitleOrTheID (also
pins SearchListRequest's scope) on all four legs — dolt, embedded dolt, uow,
served http (label-scoped, so the served leg runs it as written); handler
test that `q` reaches the seam as the search text, not TitleSearch; client
tests for `q` and the pre-dial refusal (zero dials); round-trip and
vocabulary gates; TestSearchRoutesAgree runs the same twelve flag sets on
the direct, proxied and remote routes and requires identical answers
(remote is bd serve over the proxied database, so ids and order too); served
e2e TestE2E_SearchOverHTTP. The remote legs skip with the
SearchIssuesWithCounts refusal until S6c's externaldeps reader passthrough is
underneath (loud refusal, never an empty or wrong answer); with S6c
cherry-picked all three routes agree.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* build: resync BUILD srcs and the embedded storage shard manifest for S6e/S25/S26

gazelle picks up cmd/bd/show_current.go, cmd/bd/search_remote.go, the three
backend/http e2e files and internal/httpclient/list_search_test.go; the
embedded storage shard manifest gains
TestEmbeddedReaderListQueryMatchesTheTitleOrTheID
(scripts/ci/gen_embedded_storage_shard_manifest.py 15 --weights=duration
--write).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* style: gofmt list_show_filter_modes.go (trailing blank line)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(backend/http/corpus): drop the show --current and recompute-blocked marks S6e and S25 retire

With S26 folded in, `bd show --current` resolves through
issueops.FindCurrentIssue over the Querier role and `bd recompute-blocked`
goes through storage.RecomputeBlocked (a no-op on a remote backend), so
gc-show-current and gc-recompute-blocked lose their pending marks.
mc-search stays pending, relabelled S6c+SRV-1: bd search now asks listIssues
with q behind issues.list.search, which the reference server profile does
not advertise, so the client refuses before dialing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(cmd/bd): split the bd search route agreement into the embedded and proxied lanes

TestSearchRoutesAgree skipped unless BEADS_TEST_EMBEDDED_DOLT=1, then ran
its proxied and remote legs only under BEADS_TEST_PROXIED_SERVER=1, and
its name matched neither lane's discovery (TestEmbedded* in
*_embedded_test.go, TestProxiedServer*/TestServerMode*). No CI lane ran
it, which main's #7468 guard (TestProxiedGatedTestsRunInTheProxiedLane)
now fails on.

Split it:
- TestEmbeddedSearchRoutesAgree (search_routes_embedded_test.go): the
  direct route, embedded lane.
- TestProxiedServerSearchRoutesAgree (search_routes_proxied_test.go): the
  proxied route and the remote route (bd serve over the proxied
  workspace), gated only on BEADS_TEST_PROXIED_SERVER. The remote route
  must still answer the proxied one row for row, ids and order included,
  and refuse the flag listIssues cannot carry.

Neither lane can run the other's workspace, so the shared helper
(search_routes_parity_test.go: the seed, the twelve flag sets, and
checkSearchRoute) holds every route to one expected answer per flag set;
agreeing with it is agreeing with each other. The remote leg's "needs
S6c" skip is gone: S6c is on main, so a refusal now fails.

Both manifests regenerated incrementally (embedded 100-shard block, shard
71; proxied 34-shard block, shard 23). Bazel: bd_embedded_part2_test
(--config=embedded) and bd_proxied_test (--config=doltserver-proxied)
pass and run the new tests; //scripts:go_test_sources_test passes.
Swapping the title-sort expectation fails both routes of the proxied
test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(search): keep an infra type's durable rows, expand --type aliases, fail the S26 e2e on refusal (bd-1ejjv6)

Review fixes for #7528, attempt 1: the three majors and three minors of
the synthesis.

`bd search --type <infra type>` dropped the type's durable rows (a
promoted wisp, an issue retyped after it was created), exit 0. The shared
listing (workapi.BuildListFilter) narrowed an infra --type to its
ephemeral rows unconditionally, so SearchListRequest's IncludeEphemeral
did not bring back what raw SearchIssues used to return. The narrowing
now applies only without IncludeEphemeral, in BuildListFilter and in the
count lockstep (applyCountIncludeInfra), and the http parent-walk
recognizer reads an infra type's Ephemeral back as the include_ephemeral
intent. This also fixes `bd list --type <infra type> --include-ephemeral`
and GET /v0/beads/issues, which shipped the narrowing in v1.3.1. Pinned by
workapi tests and by template, gate, wisp and promoted-message rows in the
Reader conformance Query case (new ReaderFixture.PromoteWisp) on the
dolt, embedded, uow and served backends; the issueops Reader/Counter docs
and the openapi descriptions say the same.

`bd search --type` handed the raw string to the Reader: an alias (feat,
mol, mr, ...) answered nothing and an unknown type an empty result. It is
now expanded with utils.NormalizeIssueType as `bd list` does, and an
unknown type exits 1. The help text and the CHANGELOG say so, and the
CHANGELOG also records the additive comments_omitted field in
`bd search --json` rows.

TestE2E_SearchOverHTTP skipped on the very SearchIssuesWithCounts refusal
it exists to catch; it now fails. The route agreement seeds "kinded"
template, gate, wisp and promoted-message rows and pins the alias, the
unknown-type refusal and the infra rows on the direct, proxied and remote
routes.

Minors: the CHANGELOG states the proxied `bd show --current` miss
fallback; the CommentsOmitted contract names `bd search --json`; the
storage.RecomputeBlocked and `bd recompute-blocked` comments name the
deliberate --proxied-server bypass.

Validation: go vet and gofmt on every changed package; make api-check;
workapi, issueops, httpclient/encode, types and storage unit tests; the
Reader conformance on dolt, embedded, uow and served; the embedded and
proxied route agreements (remote included), the embedded search tests,
TestProxiedServerSearch, TestProxiedServerRecomputeBlocked, the http e2e
search/show/recompute tests; docs flag, freshness and docsync checks; and
a hand-run bd search/list/count against a scratch workspace.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-13 04:12:54 +00:00
Julian Knutsen
a2d43e6792
fix(mol): molecule reads, auto-close and close --continue over http through issueops entries (S16) (#7513)
* fix(create): let a remote backend mint the --parent child id

`bd create --parent` always called the raw GetNextChildID first, which the
http backend does not serve, so every child create on a remote workspace
refused before writing anything. The Lifecycle role already mints
<parent>.<n> for a create that carries a ParentID and no id of its own, on
every backend and inside the create's own transaction; conformance
RunIssueOperationsCreateUnderAParentMintsTheNextChildID pins it on every leg,
the served one included (and the bd-enterprise postgres leg the MC server
runs). So the create route needs no new capability token: on a registered
Remote backend the client skips the raw reservation and lets the server mint.
The local route keeps its reservation so its Dolt commit staging is
unchanged, and an explicit --id beside --parent is still refused up front.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(mol): molecule reads, auto-close and close --continue over issueops roles

Over the http backend three molecule paths failed, and two of them
silently:

- `bd mol current` / `bd mol progress` without an id answered "no molecules"
  because the in_progress listing (a raw SearchIssues shape) and the parent
  walk (raw GetDependencyRecordsForIssues) were refused and the refusal was
  swallowed; with an id they refused on raw GetMoleculeProgress.
- The molecule auto-close after `bd close` never ran: findParentMolecules
  swallowed the same refused edge read into "not part of a molecule".
- `bd close --continue` claimed the next step in a raw RunInTransaction.

The direct route now reads molecules through roleMolStore, which answers
through roles every backend serves: EdgeReader (batched at the wire's 100
anchors) for the parent walk, BatchGetter for the roots (Get per id where a
server does not publish issues.batchGet), Relations for the step rows (full
rows; the raw dependents read over http is getIssue's shallow projection,
which zeroed step assignees) and the progress counts, and Reader.List for the
in_progress/hooked listings. The auto-close closes the root through
Lifecycle.Close, and the --continue claim is one Lifecycle.Update guarded by
ExpectedStatus=open.

No refusal is swallowed any more: findParentMolecules, the molecule finders
and the hooked scan return their errors; `mol current/progress` report them;
the auto-close stays best effort but warns on stderr; AdvanceToNextStep only
moves on to the next ready step when the claim LOST a race (status mismatch,
not claimable, already claimed) and returns anything else. The one heuristic
still skipped is the id-pattern heal in loadDescendants, and only for an
ErrUnsupported backend that cannot express an id-prefix search; other
failures now surface. `mol current`/`mol progress` leave the
types.IssueFilter forbidigo hole, since they no longer build a filter.

gc does not depend on bd's molecule auto-close: it runs its own
(cmd/gc/molecule_autoclose.go, from the bd on_close hook and the
controller's bead-close events). Its worker prompts do run `bd mol current`.

Served e2e (TestE2E_MoleculesOverHTTP): create --parent mints <root>.1/.2,
mol current <id> is non-empty with the full in-progress step, mol progress
counts, close --continue auto-advances, closing the last step auto-closes
the root. `mol current` without an id now refuses out loud on this stack
(asserted: never an empty list); its non-empty assertion is skipped until
S6c (externaldeps reader passthrough) is underneath.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(mol): list molecule candidates with the four wire knobs, drop the S6c skip

On top of S6c the in_progress listing behind 'bd mol current' (no id)
reaches the v0 wire encoder, which refuses ListRequest.IncludeAllTypes.
Ask for templates, gates, infra and ephemeral rows explicitly, which the
wire publishes, and run the 'mol current without an id' served case.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(issueops): one library entry each for molecule auto-close, advance and reads

The molecule rules lived only in cmd/bd, in copies that disagreed: two
auto-close implementations that closed the root through different
library entries outside the step close's transaction, a `close
--continue` claim that set no assignee and could take a step another
agent held on the direct and http routes, and three progress counts.
Every molecule rule now has one implementation in the public issueops
package, composed over three roles every backend serves (BatchGetter,
Relations, EdgeReader): MoleculeRoots, LoadMolecule, AnalyzeMolecule,
ViewMolecule, ReadMoleculeProgress, ActiveMoleculeIDs,
CompletedMolecule. Membership is the parent-child edge.

Auto-close: CloseRequest.AutoCloseMolecule and
CloseBatchRequest.AutoCloseMolecule (opt-in, so gc's own auto-close is
not doubled) close the molecule root a close completed IN THE SAME
TRANSACTION, unforced and guarded on the root revision read there:
storageissueops.CloseCompletedMoleculeInTx on the store backends (the
role bodies bound to the caller's tx, TxMoleculeReader) and
uow.CloseCompletedMoleculeInUOW on the unit of work. It replays on an
idempotent re-close, reports the root on CloseResult/CloseOutcome
.AutoClosedMolecule, and a root its own close policy refuses stays open
and is reported (MoleculeAutoCloseRefusal) without failing the step.

Advance: a new MoleculeStepper role (storage.DoltStorage accessor,
uow.MoleculeStepperSource) whose one implementation,
issueops.NewMoleculeStepper, claims the next ready step through the
Claimer (Lifecycle.Update's Claim for a wisp step): assignee set, a step
another actor holds skipped, never taken.

HTTP: closeIssue/batchCloseIssues carry auto_close_molecule behind the
issues.close.autoCloseMolecule behavior token, and POST
/v0/beads/issues/{id}:advanceMolecule (issues.advanceMolecule) serves
the stepper. Handlers only decode, call the role and encode. The client
sends the member and dials the operation only where advertised; against
an older server it DEGRADES rather than refuses — it runs the same
library rule over its own roles (issueops.CloseCompletedMolecule after
the close, NewMoleculeStepper for the advance), so the outcome is the
same in more round trips. The capability fallbacks that lived in
cmd/bd move into the client: BatchGetter answers one getIssue per id
without issues.batchGet, and EdgeReader splits past the operation's 100
anchors.

Conformance: RunMolecule* on dolt, embedded, uow and http (the http leg
twice: served, and against a server with the tokens struck out).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(cmd/bd): molecule commands and create --parent through the library entries

`bd close`, `bd close --continue`, `bd mol current`, `bd mol progress`,
`bd ready --mol` and `bd create --parent` now call the one library entry
for each rule on every route (direct, proxied server, remote):

- close (batch, --if-revision, proxied) sets CloseRequest/
  CloseBatchRequest.AutoCloseMolecule and only reports what the library
  closed in the step close's transaction. autoCloseCompletedMolecule,
  autoCloseProxiedCompletedMolecule and closeCompletedMoleculeRoot are
  gone, and so is the separate post-close write.
- --continue is the route's MoleculeStepper role: a real claim (the
  assignee is set) that skips a step another actor holds. Before, the
  direct and http routes flipped the status only and could take it.
  AdvanceToNextStep, ClaimStepIfOpen and its two implementations are
  gone.
- mol current / mol progress / ready --mol read through
  issueops.ViewMolecule, ReadMoleculeProgress, ActiveMoleculeIDs and
  LoadMolecule/AnalyzeMolecule over the route's roles. Over http this
  fixes the missing step assignee on `ready --mol` (corpus mark
  mc-ready-mol). The proxied route's SearchIssues fallback, the role
  store, the third progress count and the cmd/bd copy of the readiness
  analysis are deleted; only presentation stays here.
- create --parent asks the create role to mint the child id on every
  route (the local CLI reservation is gone) and to inherit the parent's
  labels (InheritLabelsFromParent) inside the create. The dry-run
  preview still shows the labels, and a failed label read there is
  reported, not swallowed.

TestProxiedServerMoleculeLifecycleAgreesOnEveryRoute runs one molecule
lifecycle through the real binary on the direct, proxied and remote
(`bd connect` to `bd serve`) routes and requires the same rows and the
same JSON on all three. The served e2e now also pins the held-step skip,
the claimed step's assignee and the full rows of `ready --mol`.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(httpclient): refuse molecule auto-close and advance before dialing an older server

The client used to DEGRADE against a server without
issues.close.autoCloseMolecule / issues.advanceMolecule: it closed the step
without the member and ran a second, client-side copy of the auto-close
(issueops.CloseCompletedMolecule) in its own transactions, and composed the
stepper over its own roles. That kept two executions of each rule alive.
Now the rule runs in ONE place, the server's library: a close or batch close
that sets AutoCloseMolecule, and an Advance, refuse before anything is dialed
with the typed capability error (storage.ErrUnsupported, Capability naming the
token). The composed single-item batch leg now carries the member too.

- issueops.CloseCompletedMolecule (the role-composed form) is gone.
  MoleculeRootClose is the one root close request (unforced, the auto-close
  reason, the session, guarded on the root revision read), and both
  in-transaction bodies (storage issueops CloseCompletedMoleculeInTx, uow
  CloseCompletedMoleculeInUOW) build from it.
- MoleculeView.Ready is `bd ready --mol`'s list (every analysis-ready node,
  hydrated), so ready --mol and mol current derive from ViewMolecule alone.

Tests: unit pre-dial refusal and member pins (httpclient molecule_test.go),
TestServedMoleculeRefusesAgainstAnOlderServer (step stays open, nothing
claimed) replacing the degraded contract arm.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(cmd/bd): ask for the molecule auto-close only where the route serves it

The refuse-before-dialing client made every `bd close` over http fail
against a server that does not advertise issues.close.autoCloseMolecule
(an older bd serve, or the MC server shape, whose orchestrator owns the
molecule lifecycle): the CLI always set AutoCloseMolecule.
storage.ServesMoleculeAutoClose answers, through every decorator, whether a
store's closes may carry the flag (MoleculeAutoCloseProber, implemented by
httpclient.Store from its handshake; every local backend serves it). The
direct batch close and the --if-revision close set the flag from it; the
proxied route always serves it. Against a silent server the close goes out
without the member and the server's side owns the auto-close.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(backend/http): scope the advertised-policy check to reads before the close

The close policy check is the client-side policy's last chance: it runs
before the close request, so only the reads up to that request count. A
read after the close (the old-server molecule notice reads the closed
step's edges) is not the external-dependency policy.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(backend/http/corpus): drop the six marks S16 and S19 retire

With the molecule library entries underneath, mol current (with and
without an id, json and text), mol progress and ready --mol read through
issueops.LoadMolecule over the roles, and create --parent lets the server
mint the child id. Those six shapes pass on the reference profile.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(cmd/bd): never silent against a server without the molecule tokens

Against a server that advertises neither issues.close.autoCloseMolecule
nor issues.advanceMolecule (an older bd serve, or one whose orchestrator
owns the molecule lifecycle):

- `bd close --continue` closed the step and only then found the advance
  refused, leaving the step closed and the next step unclaimed. It now
  asks storage.ServesMoleculeAdvance (a MoleculeAdvanceProber the http
  store answers from its handshake; local backends always serve it)
  BEFORE the close and refuses up front, saying nothing was closed.
- `bd close` went out without auto_close_molecule and said nothing. When
  the closed step belongs to a molecule (issueops.MoleculeRoot), a notice
  on stderr now names the molecule and the missing token: its auto-close
  is the server's to run. The batch and --if-revision closes both say it.

The served e2e masks both tokens out of the handshake and checks the
up-front refusal (the step stays open) and the notice; the httpclient
unit test pins ServesMoleculeAdvance to the handshake.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* refactor(cmd/bd): mol show through issueops.LoadMolecule on every route

`bd mol show` still loaded the molecule through cmd/bd's
loadTemplateSubgraph (raw GetIssue, GetDependentsWithMetadata and
GetDependencyRecords, plus an id-pattern child probe), while mol current,
mol progress and ready --mol read through the library. It now loads
through issueops.LoadMolecule over the route's molecule roles (the direct
store's, or the proxied provider's), so membership is the parent-child
edge everywhere and the http route reads through roles the server serves.
The served e2e checks mol show's root and step count.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs: the molecule auto-close fails the step's close on a non-policy error; stale comments

CloseRequest.AutoCloseMolecule and CloseCompletedMoleculeInTx now say what
changed when the auto-close moved into the close's transaction: a root
close the close policy refuses still leaves the root open and is reported,
but any other failure (a storage error, a lost race on the root's revision)
fails the step's close with it, where the CLI used to close the root
afterwards and only warn.

Stale: the advanceMolecule vocabulary entry said the client composes the
stepper over its own roles when unserved (it refuses before dialing), and
the graph-create collision probe called itself best-effort by transport
(the http GetIssue miss is storage.ErrNotFound like every other store's).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(changelog): the molecule tokens and the S16/S19 behaviours

Entries for issues.close.autoCloseMolecule and issues.advanceMolecule, the
issueops molecule entries, the auto-close inside the step's close (and a
non-policy failure of it failing the close), the real claim behind
close --continue, the molecule reads on every route, the notice and the
up-front --continue refusal against a server without the tokens, and
create --parent / mol reads over http.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(httpclient): regenerate unsupported_gen.go for MoleculeStepper

The PR adds MoleculeStepper() to storage.DoltStorage (200 -> 201
methods). httpclient implements it by hand, but the generated shell's
trailer still counted 200, so
//internal/storage/unsupportedgen:unsupportedgen_test
(TestGeneratedShellIsFresh) failed in Bazel / test and
Bazel / integration, turning CI Gate / Required red.

Regenerated with `go generate ./internal/httpclient`; the only change
is the trailer count (143 of 201 methods generated; 58 left).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(molecule): auto-close refusals, write skew, close --json, docs

Review fixes for #7513, numbered as in each iteration's review
synthesis.

Iteration 1:

F1: a close guard's refusal of the root (TemplateReadOnlyError,
PinnedError, CloseNotAssigneeError) is an auto-close refusal, like the
close policy's: the root stays open and is reported, and the step's
close stands. Pinned by the new conformance case
RunMoleculeGuardedRootKeepsTheStepClose on dolt, embedded, uow and the
served http client.

F2: two closes that finish a molecule's last two steps in concurrent
transactions each read the other's step as open, so neither closed the
root and both committed. Every auto-close now rewrites the root's
coordination cell before it decides the root (local_metadata
molecule-coordination/v1/<shard>, the dependency coordination idiom and
sharding with a FreshRowLock value), in both twins
(CloseCompletedMoleculesInTx and CloseCompletedMoleculesInUOW). Two
deciders of one root now conflict at commit (1213), and the retry reads
the winner's close and closes the root. This is the synthesis's option
(c), applied through a coordination cell instead of the root row; it
needs no post-commit second transaction and no engine lock. The
deterministic two-transaction regression fails without the touch and
passes with it on dolt and uow; the concurrent-close tests end with the
root closed in every round, reported by exactly one close.

F3: the auto-close checks the root's eligibility before walking any
step, MoleculeStepsClosed stops at the first open step, and a batch
decides each root once and reports it on the first item under it.

F4: ErrVersionMismatch on the root is not a refusal; it fails the
step's close with it. The classifier unit test pins this. A conformance
case cannot stage that race deterministically through the public roles.

F5: storage.ServesMoleculeAdvance answers (bool, error), and bd close
--continue refuses before closing anything when the handshake cannot be
read.

F6: the CHANGELOG and the stepper and batch-closer doc comments say
what the code does.

F7: the unserved-auto-close notice fires only when the close left an
auto-closing molecule with every step closed and its root open
(issueops.CompletedMolecule): again on a re-close, once per root in a
batch, never for a step under an ordinary epic or an unfinished
molecule. Proxied mol show, mol current and mol progress again report a
read unit of work that cannot be opened as itself, not as a molecule
that was not found.

F8: the write-ledger comment says the close is refused before dialing.
The dead FirstClosed/LastClosed guards in mol progress pre-exist on main
and are left as they are.

Iteration 2:

F1: on the proxied route, bd ready --mol reported a read unit
of work it could not open twice: once as itself, then as "molecule
'<id>' not found", which under --json is a second error document on
stdout. It now returns moleculeNotResolved, as mol show, mol current and
mol progress do. TestProxiedReadyMolReportsAnUnopenableReadOnce drives
runMoleculeReady with a provider whose NewUOW fails and wants one Error:
line, or one JSON document; both subtests fail with the old return.

F2: the coordination cell serializes every concurrent close of a step
under an open auto-closing root, not only the closes that finish the
molecule. That is the price of deciding the root in the step's
transaction, so it is documented (TouchMoleculeCoordinationInTx, the
UOW twin, CloseRequest.AutoCloseMolecule, the CHANGELOG) and pinned on
dolt and uow: two sibling closes with a third step still open conflict
(both commit when the touch is removed, on either twin), and a six-step
fan-out closes every step through the store's retry, with the root
closed and reported by exactly one close. The synthesis's alternative,
a post-commit recheck by any close that read a step open, is not taken:
it is a second, best-effort transaction after the step's commit, the
shape this PR exists to remove.

F3: MoleculeRootClose's doc no longer promises that a step reopened
between the completion read and the root close fails the step's close.
The guard is the root's RowVersion, which a step reopen never writes;
the doc now says what the guard covers and that this race goes
undetected (it is inherited from the CLI's old post-commit auto-close).

F4: the user docs describe the auto-close. docs/workflows/molecules.md
replaces "Closing the last child does not close the molecule root"
with which roots close and which stay open, the refusal warning, the
unserved-server note and the --json members, and documents the claiming
--continue and parent-child step membership;
docs/reference/json-schema.md gains a bd close --json section. These
are main-only behaviors on pages pinned to v1.2.2 (docs/cli-docs.pin):
a maintainer call against §1 of
engdocs/decisions/2026-07-17-docs-release-pin.md, as in #6508, #5855,
#5997 and #5092. The auto-close itself shipped in v1.2.2
(cmd/bd/close.go autoCloseCompletedMolecule), so the old sentence was
already wrong at the pin. docs/cli-reference/ is generated from the
pinned release and is not hand-edited; the bd close and bd mol show
help texts carry the same facts and reach it at the next pin bump.

F5: bd close --json reports the auto-close. The entry of a close that
completed a molecule gains auto_closed_molecule (the root it closed) or
molecule_auto_close_refusal (why the root stayed open), the HTTP API's
names, on the direct, proxied and --if-revision routes. The members sit
beside the issue's own rather than wrapping the output in an object the
way --suggest-next does: an object would change the output's shape
exactly when a close completes a molecule and break every consumer that
reads an array, while members leave each entry decoding as the closed
issue, compose with `closed` under the flag objects, and leave a plain
close's bytes unchanged. Pinned on the embedded and proxied tiers
(TestEmbeddedCloseJSONReportsMoleculeAutoClose, TestProxiedServerClose2B
json_reports_*, and both --if-revision replay tests), all red without
the change.

F6: cmd/bd's BeadsTemplateLabel is issueops.MoleculeTemplateLabel, so
the template commands and molecule membership cannot drift apart.

F7: the UOW twin's revision probe says that it writes nothing,
and the BatchCloseRequest.auto_close_molecule spec description has its
verb back (apigen regenerated, no drift).

Iteration 3:

F1: on the unit-of-work route (bd serve, the proxied CLI), a refused
root's molecule_auto_close_refusal and warning carried the use case's
and repository's wrapping ("close <id>: db:
IssueSQLRepository.CloseChecked <id>: ...") around the refusal, where
the tx twin and the docs give the refusal alone. Both twins now report
it through RefusedMoleculeAutoClose, which answers the refusing guard's
or policy's own sentence. RunMoleculeRefusedRootKeepsTheStepClose,
RunMoleculeGuardedRootKeepsTheStepClose and the embedded and proxied
close --json tests want that exact sentence; the uow conformance leg
and the proxied test fail without the unwrap.

F2: the docs no longer promise ErrVersionMismatch for a write to the
root that races its auto-close. The root's read and its revision check
share one transaction snapshot, so that write conflicts with the root
close at commit and the retry decides the root again; the close fails
only if it still conflicts when its retry budget runs out
(MoleculeRootClose, IsMoleculeAutoCloseRefusal,
CloseRequest.AutoCloseMolecule, the tx twin's comment, the CHANGELOG
and the OpenAPI description, apigen regenerated).

F3: TestEmbeddedCloseJSONReportsMoleculeAutoClose moves above
TestEmbeddedCloseConcurrent's doc comment, which documents its own test
again.

F4: bd mol progress counting direct steps while the auto-close counts
every descendant predates this PR; it is left for a follow-up.

F5: cmd/bd's molReader drops GetMoleculeProgress, which nothing called
once mol progress read through issueops.ReadMoleculeProgress.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-12 19:50:34 +00:00
Stephanie Jarmak
559b5de1e2
fix(memory): preserve presence for empty values (#5964)
* fix(memory): preserve presence for empty values (#5963)

Recall and Forget now distinguish an existing empty memory row from an absent key across direct, embedded, and UOW storage paths.

Agent-Signature: codex-unknown-model-unknown-reasoning on behalf of sjarmak

* fix(memory): make bd recall report presence from row existence

printRecallResult (and the recallCmd caller) derived "found" from
value != "", so a memory stored as the empty string still printed
"No memory with key" and exited 1 even though the role's
RecallResult.Found was true and bd forget would delete the row.
Thread result.Found through instead: an empty-valued memory now
recalls cleanly (exit 0, empty value; JSON reports found: true), and
only an absent key is a miss.

Addresses maintainer review on #5964 (fixes #5963).

* test(httpapi): pin GET /v0/beads/memories/{key} as 200 for an empty value

TestGetMemoryAnswersAMissWithA404 pinned a RecallResult{Value: ""}
case (Found defaulting to its zero value, false) as a 404, a shape no
backend can produce since Found now comes from row existence rather
than value truthiness. Move that case into its own
TestGetMemoryAnswersAnEmptyValueWith200, which recalls Found: true
and asserts a 200 with an empty value. Also updates the handler's
doc comment, which still described the retired 404-on-empty
behavior.

Addresses maintainer review on #5964 (fixes #5963).

* docs(httpapi): fix openapi spec to say memory presence is row existence

The published spec still said an empty-valued memory is a 404 on
GET /v0/beads/memories/{key} and DELETE /v0/beads/memories/{key},
contradicting the role's RecallResult.Found (row existence). Rewrite
the operation descriptions, both 404 response bodies, Memory.value
and RememberedMemory.replaced to say presence is row existence
everywhere: an empty-string memory is a 200 (or is removed by
forget) with an empty value, matching what GET /v0/beads/memories
already enumerated. Regenerate types.gen.go via `make api-gen` so
the generated doc comments match.

Addresses maintainer review on #5964 (fixes #5963).

* docs(changelog): note the empty-valued memory presence fix

Add an Unreleased/Fixed entry for #5964 (fixes #5963): bd recall,
bd forget and GET /v0/beads/memories/{key} now treat an empty-valued
memory as present everywhere, instead of only on GET
/v0/beads/memories.

* fix(memory): make bd remember <bare-key> recall an empty-valued memory

rememberBareKeyPath decided presence from existing != "", so a bare slug
naming a memory stored as the empty string was refused as "no memory
named" instead of recalled. It now takes the role's Found, matching
printRecallResult, bd forget and the HTTP door.

* test(memory): pin empty-value presence through the real commands

TestPrintRecallResult and TestRememberBareKeyPath call the helpers directly,
so a RunE that went back to deciding presence from value != "" would leave
them green. TestEmbeddedMemory/empty_valued_memory_is_present empties a
memory row beneath bd and runs bd recall, bd remember <bare-key> and
bd forget against it, checking first that bd no longer sees the placeholder.

Two comments in problem.go and one in memories_test.go still said an
empty-valued memory is a 404. The CHANGELOG line now says a key with no row
is a miss, since a forgotten key was once stored.

* build(bazel): add the memoryops dep to the bd_test go_test (make bazel-sync)

memory_embedded_test.go now imports internal/storage/memoryops directly
(storagememoryops.StorageKey, to overwrite a memory row out of band), so
gazelle emits //internal/storage/memoryops on bd_test. cmd/bd/BUILD.bazel
landed on main after this PR branched, so the PR could not carry the dep;
without it, bazel.yml's "BUILD files in sync" step fails in the required
bazel-test lane. Generated by `make bazel-sync`.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(memory): maintainer fixup — adoption review nits

Maintainer fixup for the adopt-pr iteration-1 review of #5964. Both review
lanes approved and the quality scorecard approved at 981; these are its four
non-gating nits, each confined to lines this PR already adds.

- internal/httpapi: TestForgetMemoryOfAnEmptyValuedKeyIs200 pins the DELETE
  door's empty-value 200 the way TestGetMemoryAnswersAnEmptyValueWith200
  pins the read. Guarding the handler's 404 on Value == "" reddens exactly
  this test and nothing else.
- CHANGELOG: the bullet names DELETE /v0/beads/memories/{key} beside GET,
  whose answer changed the same way (404 with the row left standing -> 200
  with the row removed).
- uow: the two new GetAllConfig presence probes in Recall and Forget say why
  they read the whole table, by pointing at Remember's rationale.
- conformance: RunMemoriesEmptyRowIsPresentUntilForgotten runs its
  no-history-on-retry leg last and SKIPs it with a reason when a fixture has
  no CountHistory, as the MemoriesFixture doc promises, instead of passing
  quietly. All three in-tree fixtures wire the hook, so every in-tree run
  is unchanged (19/19 on dolt, embeddeddolt and uow, 0 SKIP).

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(httpclient): report an empty-valued memory as present (base-refresh repair)

Maintainer repair from refreshing #5964 onto main b630f96f5b. Main's #7247
added the HTTP client store (internal/httpclient) after this PR branched,
and its memory role and served conformance leg still treated a memory
stored as the empty string as absent:

- Recall set Found from Value != "".
- The RecallMemory doc said an empty-valued row was a 404.
- TestServedMemoriesRecallConflatesStoredEmptyWithAbsent called
  conformance.RunMemoriesRecallConflatesStoredEmptyWithAbsent, which this
  PR deletes.

The rebase applied cleanly, but the merged tree did not compile.

- internal/httpclient/memories.go: Recall reports Found true on any 200.
  With this PR the wire returns a 404 only for a key with no row; an
  empty-valued row comes back as a 200 with an empty value. Setting Found
  from Value would hide that row again on the HTTP backend. Forget already
  reported Found true on a 200.
- internal/httpclient/wire/writes.go: the RecallMemory doc now says the
  same.
- internal/httpclient/served_memories_deps_test.go: the hm10 served case
  now runs RunMemoriesEmptyRowIsPresentUntilForgotten, the contract case
  this PR adds in place of the one it deletes.

With BEADS_TEST_EMBEDDED_DOLT=1 the served memory tests pass with 0 SKIP.
Reverting the Recall line makes hm10 fail with
`Recall("hm10-stored-empty") = value "" found false, want value "" found
true`.

Agent-Signature: claude-code-claude-opus-5-5-max on behalf of Gas City Adopt-PR
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: sjarmak <t@t.co>
Co-authored-by: Eddie the Engineer <adopt-pr@gascity.local>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-12 19:28:54 +00:00
Julian Knutsen
b630f96f5b
fix(httpclient): unlimited ready/query/list reads over http, client-capped (S15) (#7497)
* fix(httpclient): unlimited ready/query/list reads over http, client-capped

An unlimited read never puts limit=0 on the wire, which a non-loopback
bd serve refuses ("unlimited reads are loopback-only"):

- bd ready --limit 0 (the WorkFilter bridge) sent no limit at all, so the
  server's 100-row ready default silently capped every routed-pool queue.
- Reader.Ready and Querier.Query with Limit 0 sent limit=0 and got a 400
  from a non-loopback server.

listReadyWork and queryIssues publish no cursor, so an unlimited read is
one request bounded at the client cap plus one (one server read, so one
snapshot). listIssues keeps its cursor walk, now bounded by the same cap
(per-page reads; the snapshot semantics are documented in
unlimited_read.go). Past BEADS_HTTP_UNLIMITED_READ_CAP (default 10000) the
read is refused with *UnlimitedReadCapError, never truncated; a server that
reports has_more below the requested bound is refused the same way. A
dispatch tripwire refuses limit=0 on the three listing operations, so a new
path that forgets fails loudly instead of only off loopback.

Served tests run against a server in --allow-non-loopback mode with 150
rows: ready (role, bridge, bridge+total), query and list at limit 0 return
every row with a bounded limit on every request, and the cap refuses with
the typed error.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(backend/http/corpus): drop the three marks S15 retires

With unlimited reads served (one request bounded at the cap plus one for
ready and query), gc-ready-unlimited, gc-ready-pool-demand-unlimited and
gc-query-wisps-unlimited return every row on the reference profile instead
of the server's default page or the loopback-only refusal. mc-list-id stays
pending on S15b (--id).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(changelog): unlimited reads over http and BEADS_HTTP_UNLIMITED_READ_CAP

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(httpclient): uncap the list walk and bound the ready-count fallback

Review fixes for #7497 (review loop iterations 1 and 2).

F1 (major): the client cap no longer applies to an unlimited list.
listIssues publishes a keyset cursor, so no response is larger than a
page; the walk is back to base, bounded only by --max-rows /
BEADS_MAX_ROWS as on every local backend. The cap stays on the
cursorless reads (ready, query), whose whole answer is one response the
server buffers. The CHANGELOG, unlimited_read.go and the env doc say so.

F2 (major): externaldeps.CountReadyWork's no-counter fallback, which an
http store reaches whenever its server does not apply the
external-dependency policy (every released server), asked for the whole
ready set to report one number: over http, up to the cap's full rows,
and a refusal past it that text `bd ready` swallows into a missing
footer. It now asks for a one-row Lite page; the bridge pairs it with
the server's countReadyWork, and refetchedReadyTotal still sizes the
externally blocked case.

F3: TestCheckUnlimitedPage pins a full cap+1 page that also reports
has_more as the cap's refusal, not a truncation, and the env test
accepts the inclusive 1000000 bound. Served: the cap test adds the
unlimitedRows-2 case (the server's 149 rows with has_more), the list is
pinned as uncapped with MaxRows its only bound, and the bridge's
MaxRows-vs-cap precedence is pinned with the limit on the wire.

F5: readyBridgePage sends MaxRows+1 rather than cap+1 when MaxRows is
the tighter bound (tightenUnlimitedPage). The MaxRows check still runs
first, so an answer past it keeps its *ErrTooManyRows attribution.

F6: the decorator and test comments describe the cap+1 bound; the
refusal reads "narrow the query or raise BEADS_HTTP_UNLIMITED_READ_CAP";
the Ready and Query Limit contracts name the http exception; mc-list-id's
pending reason drops the S15 --limit 0 clause and keeps the --id refusal.

F7: unlimited_read.go states that the tripwire guards Store.dispatch
only, so a direct wire.Client caller (readyClaimPage) must send a
positive limit itself.

Not taken: F4 (an unlimited walk that meets has_more with no cursor ends
as a success rather than a ServerTruncated refusal). After F1 the walk
is base code again, the shape needs a skewed or older peer, and the
change would land in fetchIssuePages (cyclomatic 21, cognitive 34),
turning an untouched base function into a changed one in the major
complexity band. Left as a follow-up.

Iteration 2, N1: F5 reached only the raw bridge. `bd ready` reads
through the external-dependency decorator, and against a server that
leaves that policy to the client, one externally blocked issue made
readyExclusion take MaxRows off the unlimited inner query. The bridge
then asked for cap+1 rows: past the cap it refused with the read cap's
error (exit 1, "raise BEADS_HTTP_UNLIMITED_READ_CAP") rather than
--max-rows' (exit 2), and under it fetched the whole ready set to refuse
on the rows kept. An unlimited inner query now keeps MaxRows plus one
row per blocked issue: past that many rows more than MaxRows are kept
whatever the drop takes, so the bridge sizes the request by it under the
cap, and innerCapError hands the refusal back as the caller's own cap.
The suggested inner Limit of MaxRows+blocked+1 was not taken: a bounded
request skips the cap and its server-truncation check, so with any
external blocker a --max-rows above the cap would fetch past it.
TestServedReadyBridgeMaxRowsSizesTheRequest pins the request and the
refusal through externaldeps.New with one blocker, the cap still in
force past MaxRows plus the blocked rows;
TestServedReadyMaxRowsCountsTheRowsKept adds an unlimited case whose
refusal names the caller's cap on all three decorator reads.

N4: QueryRequest.Limit's "NOTHING BACKSTOPS IT" names the http
exception in its lead, as ReadyRequest.Limit does.

Not taken in iteration 2, as non-gating follow-ups: N2
(refetchedReadyTotal's Limit = total+blocked read sits outside the cap;
pre-existing), N3 (the list walk's has_more-without-cursor end, the
deferred F4), N5 (prefixing the malformed-cap error with its operation;
it already names the variable and its value).

Validation: the served tier (BEADS_TEST_EMBEDDED_DOLT=1
BEADS_HTTP_TEST_REQUIRED=1) for internal/httpclient and the command
corpus; go test for externaldeps, issueops and workapi. Each new test
was mutation-checked red: the zero-limit count, the tightening removed,
the MaxRows check removed, the PR's list cap restored, and a cap+1
has_more page read as a truncation. Iteration 2: the served
internal/httpclient ready, unlimited-read and external-policy tests (34
PASS, 0 SKIP) and go test for externaldeps, issueops and workapi; the
new pins go red with MaxRows taken off again (the pre-fix code), with no
re-attribution, with no growth by the blocked rows, with the suggested
inner Limit, and with Found left unadjusted.

Agent-Signature: claude-code-claude-opus-5-5-unknown-reasoning on behalf of Julian Knutsen
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-12 09:02:22 +00:00
Julian Knutsen
22c0faaaf9
test: fix the three causes of Main's Test (macOS) failures (#7494)
* test(dbproxy/server): disable dolt event flush so send-metrics cannot race TempDir cleanup

Every dolt sql-server this suite stops cleanly forks a detached
`dolt send-metrics` child on exit. It outlives the test, rewrites
<HOME>/.dolt/eventsData (HOME is a t.TempDir()), and races t.TempDir's
RemoveAll: "TempDir RemoveAll cleanup: ... 001/.dolt: directory not empty".
main.yml's Test (macOS) job (plain go test) failed on this in about half
of the pushes since 2026-10-09; Bazel never did because
tools/bazel/test_env.sh exports DOLT_DISABLE_EVENT_FLUSH=1. TestMain now
exports it too, guarded by TestSuiteDisablesDoltEventFlush.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* test(scripts): skip Linux-CI-only rrc and worker-env-sync script tests off Linux

TestRRCRunLanes runs .github/scripts/rrc-run-lanes.sh, which uses bash 4's
`mapfile` (exit 127 under macOS /bin/bash 3.2). TestWorkerEnvSyncCopyAgainstFixture
runs tools/rbe/worker-env-sync-copy, whose GNU `sed -i -E` becomes
"-i with backup suffix -E" plus a basic regex under BSD sed, so the pin is
never rewritten. Both scripts run only on Linux runners (bazel.yml
rrc-seed/rrc-verify, worker-env-autosync.yml); per their owner they stay
Linux-only and the tests skip elsewhere with the reason.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-11 02:31:35 +00:00