* fix(metadata): restore v1.2.2 scalar typing for --set-metadata values
8ebd53e88 moved the CLI's metadata-edit helpers into
internal/storage/metadata.go and, in the move, replaced v1.2.2's
toJSONValue inference with a MetadataEditValue that always json.Marshals
the value as a string. The intent was GH#4146 (a numeric-looking id or a
version like "1e3" came back as a JSON number, breaking map[string]string
Go consumers), but it silently changed the shipped CLI contract:
`bd update --set-metadata k=5` started storing "5" instead of 5. That
regression shipped only in the deprecated/superseded v1.2.1, never in
v1.2.2, and 1.3.0 would freeze it.
Restore v1.2.2's inference body-for-body inside MetadataEditValue, so the
parity claim stays checkable against `git show v1.2.2:cmd/bd/update.go`:
"null" is JSON null, "true"/"false" are booleans, a spelling that scans
as a float AND is valid JSON on its own is emitted as that literal
(5 -> 5, 1e3 -> 1e3, 0.50 -> 0.50, while 05, +5 and NaN stay strings),
and everything else is a string. Every --set-metadata caller funnels
through this one function: embedded (cmd/bd/update.go ->
storage.ApplyMetadataEdits), proxied (parseSetMetadataFlags) and the
update executor's []string path.
GH#4146's concern is served by the explicitly typed routes, which are
untouched: --metadata / --metadata-json on the CLI (MergeMetadataJSON)
and the typed set map on the HTTP/Go path (applyTypedMetadataEdits).
No data migration: a "5" written by an in-window build is
indistinguishable from a string a user set intentionally through the
typed paths, so normalizing stored metadata would corrupt intended data.
Re-running the same --set-metadata k=5 retypes any given key.
Tests: new table in internal/storage/metadata_edit_value_test.go pinning
the v1.2.2 matrix plus an ApplyMetadataEdits end-to-end; the tests that
pinned the string-always behavior are flipped (domain/db mergeops suite,
update_lifecycle patch mapping, cmd/bd toJSONValue table) and the
embedded + proxied CLI suites now assert a stored JSON number. The
contract corpus is regenerated: its update/list/ready goldens run
`--set-metadata phase=2`, so `"phase": "2"` becomes `"phase": 2` and
nothing else moves.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(create): echo full-precision timestamps from bd create, not the DATETIME(0) re-read
b92442d1a routed `bd create` through issueops.Lifecycle, whose executors
build their result by re-reading the row they just wrote
(HydrateIssueOperationResult -> GetIssueInTx). created_at/updated_at are
DATETIME(0) columns, so that snapshot comes back at whole-second
precision and cmd/bd/create.go echoes it. v1.2.2 returned the caller's
own in-memory struct, which PrepareIssueForInsert stamps with
time.Now().UTC() including nanoseconds, so `bd create --json` used to
print e.g. 2026-08-28T12:34:56.789012345Z. Agents that create several
issues in a burst order them by exactly that fraction; rounded to the
second they all tie.
Restore the echo with OverlayCreateTimestamps: after hydration, put back
the created_at/updated_at (and closed_at, when the write set one) the
INSERT actually sent. Everything else on the snapshot stays DB-truth,
which is the facade's contract — only create, and only these fields.
Applied at every create body, not just the two the CLI reaches:
issueops.ExecuteCreate and ExecuteCreateBatch behind the dolt and
embedded-dolt stores, and the unit-of-work provider's own Create,
CreateBatch and batch-apply create item, which is what the proxied server
and POST /v0/beads/issues run on. Fixing only the shared executors would
have left the two bodies disagreeing about the same promise.
What is NOT changed, because it is not a regression: `bd show --json` and
`bd list --json` answered whole seconds in v1.2.2 too — they read the
column — and update/close/reopen re-read after writing there as well.
Widening the columns to DATETIME(6) would give reads sub-second times
too, but that is a schema migration on user Dolt DBs and touches paths
that deliberately compare at second precision; it is a 1.4 change.
The accepted skew, which predates this and stays: the engine ROUNDS to
the column rather than truncating, so a later read of the same row can
sit up to half a second either side of the create echo.
Tests: two new conformance cases, run on all three backends —
RunLifecycleCreateEchoesSubSecondTimestamps (a caller-supplied
nanosecond stamp must survive to the result exactly; the auto-stamped arm
stays a bounded smoke check, since requiring a non-zero fraction off the
clock is flaky by a billionth) and RunBatchCreatorEchoesSubSecondTimestamps
(per item, since both bodies build the result one snapshot at a time).
Plus a cmd/bd embedded test asserting `bd create --json` prints a
fractional part, and one httpapi assertion that the wire does not round
the role's snapshot on the way out. All four fail on the pre-fix tree.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(httpapi,types)!: revision / expected_version tokens are opaque JSON strings
The optimistic-concurrency token is minted by freshRowLock over the FULL
int64 range (crypto-random, non-zero) and published on the wire as a JSON
number: `revision` on responses, `expected_version` on guarded requests.
A JSON number does not survive a consumer that decodes into an IEEE-754
double — JavaScript's JSON.parse, Go's `any`, Python's `float`, jq. Live
tokens run past 5e17, where a double's ulp is already 64, so such a
client reads back a value NEAR the token, echoes that as its guard, and
earns a 409 against a row nobody else touched. The spec already carried
"DECODE IT AS A 64-BIT INTEGER" paragraphs telling clients to work around
this; that is a workaround for a wire type that should not have been a
number.
Emit and accept the token as a decimal STRING instead. It is documented
as opaque and EQUALITY-ONLY — compare it, never order or interpret it —
and a string round-trips exactly in every JSON consumer. It also leaves a
future backend free to mint a token that is not an int64 at all.
Internals are untouched: freshRowLock, the row_lock BIGINT column, its
migration, types.Issue.RowVersion and the ExpectedVersion *int64 guards
on the issueops requests all stay int64. Only the wire projection moves,
through the one spelling of the encoding — types.RevisionToken and
types.ParseRevisionToken.
STRICT STRING ONLY, with no transitional number-or-string acceptance.
This API already refuses unknown members by name and keeps one spelling
of the guard, and a number is precisely the shape a double-based client
has already rounded before it left the client — accepting it would
preserve the defect. A numeric guard is a 400 naming the member.
Chosen over clamping freshRowLock to 53 bits: a clamp does not fix rows
already minted with full-range tokens by in-window builds, so a JS
consumer reading one still rounds past 2^53 — repairing that would need a
row_lock normalization sweep over issues and wisps, a migration days
before a tag.
Surface: types.IssueDetails.Revision (what `bd show --json` prints and
GET /v0/beads/issues/{id} returns), the five hand-assigned httpapi
response fields, applyVersionGuardMember (now the reader for the batch
update item too, ahead of its typed decode, so a mistyped token still
earns a 400 naming the member), Problem.expected_version, and 14 OpenAPI
members regenerated through `make api-gen`. actual_version goes with them
— it is documented as "the `revision` the row was found holding", so
leaving it an integer would make the expected/actual pair disagree about
its own type; it is declared but never populated, so nothing changes
behaviorally.
Nothing shipped in a tagged release ever carried this member: v1.2.2 has
no internal/httpapi at all and no `revision` anywhere on the wire, so the
whole surface is in-window and can still be changed compatibly. Legacy
migration-0054 rows read as the string "0" with the same CAS semantics.
Corpus goldens are UNCHANGED: the canonicalizer already replaced any
`revision` value with the string "<REVISION>" whatever its JSON type.
Tests: types-level pins that the marshaled member matches `"revision":
"-?\d+"` and that RevisionToken/ParseRevisionToken round-trip MinInt64,
MaxInt64 and 2^53+1, plus a JS-realistic decode of 2^53+1 with a float64
control proving the token really is one a number would corrupt; httpapi
refusal tests for a numeric guard on update, close, reopen, delete and
both batch item kinds; a 409 echo asserted as the string; and the proxied
serve integration tests now splice the response's own token back in
verbatim rather than re-rendering a parsed one.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Three beads (ga-uwix0.1, ga-2yaqp.5, ga-2yaqp.6) plus a fable review round.
Docs and spec prose only: all 19 touched .go files are gofmt-normalized,
comment-stripped byte-identical to main. The one exception is deliberate and
reviewed — scripts/check_doc_freshness_test.go gains a data row, because the
delegated gate ruling required it.
ga-uwix0.1 — #5516 gave bd serve optional bearer auth and made it MANDATORY
beyond loopback; fifteen prose sites still said the opposite. The three the
council named are reworded to the stronger true form (the actor is not the
authenticated principal even where a bearer is required; a shared token admits
a client to the WHOLE surface and grants no narrower right). The sweep for a
fourth found eight more in five spellings the audit's grep could not reach,
including claim.go's "any client can claim as any name", release.go's
no-AUTHORIZATION variant, and six memory/settings-plane sites where the true
statement is sharper: a bearer decides WHO may call, never what a caller who
is let in may read. The review round found four more in the CHANGELOG's
UNRELEASED section, which round one had misclassified as released history —
they would have shipped false in the same release as the auth entry.
The serve runbook's "Binding beyond loopback" section was actionably wrong: it
said the flag "prints a warning and changes nothing else", and its refusals
table quoted a string not in the binary. Rewritten against a build, with every
flag, error string, event name and reason value produced by running the
command. It gains the operational trap nobody had written down: /healthz is
auth-exempt and GET /v0/beads/ready is not, so an unauthenticated readiness
probe earns a 401 forever while liveness stays green. The runbook is now IN
scripts/check-doc-freshness.sh, pinned to the four files it is read against.
ga-2yaqp.5 — nine prose defects, every count re-derived rather than trusted;
three of the audit's own numbers had moved again. 38 route rows and 36
capability-bearing ones; thirteen pinned schemas, not twelve or seven; six
missing entries in the no-query-parameter list, not five; twenty-eight role
accessors, not twenty-seven; nineteen patch members, not thirteen. Two prose
"code counts" were counting paragraphs. CodeAlreadyClaimed claimed the holder
is always attached when only one of four producers always attaches it and
releaseIssue never does. MetadataCAS's leaf contradicted its own shipped wire.
The path/operationId inversion is now recorded where a spec author looks.
ga-2yaqp.6 — the role checklist refreshed against what this wave's additions
actually cost, traced to real commits. Step 10 sent people to a name list
#5460 deleted. The "four more edits" claim was stated twice and undercounted
both times: the two freshest wire slices share a nine-file core plus
roles_test.go. The step-with-no-number gained the fact that explains its
instances — write-role decorators wrap, read-role decorators recurse, so the
call lands on the stub, and both real instances are reads. The wave got its
first CHANGELOG entry: six new public operations, not three, and zero new role
accessors in the window.
Review verdict: FIX-REQUIRED then all items addressed. The reviewer
live-falsified the strong auth form and failed to break it, and confirmed
every recount re-derives exact.
* Publish POST /v0/beads/issues/{id}:release: give a claim back (ga-lzcuv)
The claim has had a wire since v0 and its inverse has had none. `bd unclaim`
is a subprocess for every client of this surface, and a release-if-current has
no route at all — so the one verb that frees stranded work is the one verb an
automation client cannot reach.
`releaseIssue` publishes issueops.Releaser whole: `actor`, the
`expected_assignee` compare-and-set, and `force`. The response carries the
issue, `changed`, and `revision` — the last for `updateIssue`'s reason, and
one of its own: a release REMINTS the row's token by design, so a caller that
guarded its next write on a version read before the release would miss, and
this member is the only place the new one exists.
THE REFUSAL CODES ARE THREE, AND ONLY ONE IS NEW.
`already_claimed` is the ownership fence — an unforced, unguarded release by
an actor that is not the holder. That is `updateIssue`'s assignee-fence
situation exactly, down to having two bypasses, so it takes that operation's
code rather than a fourth spelling of "a live foreign owner refused your
write". Like there, no `assignee` member: the fence refuses without naming the
holder, so absence means re-read the row.
`precondition_failed` is the guard, inherited from the same operation and
carrying the same members for the same reason — `expected_assignee` is the
REQUEST's value, and there is no `actual_assignee` beside it, because the
refusal rolled its transaction back and ErrAssigneeMismatch is a bare sentinel
that formats the holder it saw into prose. There is no ClaimConflictError
equivalent to read, and this surface does not scrape its own messages.
`not_releasable` is minted, covering ErrNotReleasable and ErrNotClaimed.
`not_claimable` is the nearest existing code and covers a superset of the same
statuses today, but the coincidence is not a contract — the release transition
is pinned to {open, in_progress} by releasableStatus and the claim's
eligibility is a separate predicate — and for the unheld half the reuse would
be an outright lie: an open, unassigned row is the most claimable row a
workspace has. The two conditions share ONE code with no discriminating
member, which is the narrow choice rather than the convenient one: neither
sentinel carries a typed observation, and splitting later is an ADDITION the
document already tells clients to tolerate, while merging two published codes
is the removal that breaks the wire.
IT IS NOT IDEMPOTENT, and the wire says so rather than inventing an answer the
role refused to give. An unheld row is 409, never a 200 carrying
`changed: false` — that value is one no role produces (ReleaseResult.Changed
is true on every non-error answer), and it would report one answer for "I
released this twice", "a reaper beat me to it" and "nothing ever claimed it",
because the post-state is anonymous and nothing is left on the row to tell
them apart. A client for which "already released" is the ordinary path maps
the code to success in one line it writes knowingly.
`expected_assignee` disagrees with `updateIssue`'s same-named member in
exactly one place, and the schema says so: the empty string is a 400 here
rather than a guard meaning "expected unassigned". It also documents what the
comparison ACTUALLY does — separator-insensitive over runs of `.`/`_`/`-`,
per actorMatches, and nothing else forgiven — rather than repeating the role
doc's "byte for byte", which the implementation stopped being at ga-5ksp5.
failRelease matches all four sentinels TYPED and BEFORE its ErrValidation and
ErrNotFound arms. None of them is wrapped in ErrValidation on any leg, so
below a generic arm every one would have been a 500 on every leg.
The real-Dolt leg asserts the property a fake cannot: after every refusal it
reads the row back through the CLI, so "nothing was written" — and above all
the CLAIM SURVIVING a wrong-holder refusal — is a fact about the stored row.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Fix the release document's three review findings (ga-lzcuv)
The mint stands; what the review caught is that the document was written as
though `not_releasable` meant one thing, and it means two.
THE SAFETY ONE. "A caller for which 'already released' is the ordinary path
treats that code as success" was true of the sentinel it was written against
and is FALSE of the code it shipped as. `not_releasable` also covers the
status refusal, and a row parked in a status this workspace configured is
refused WHILE STILL ASSIGNED — so a reaper following that sentence books a
claim as dropped while it stands, held by an agent that is already gone, with
nothing left to free it. That is the stranding this operation resolves ids
across both planes to prevent, reintroduced through the refusal vocabulary
instead of through the id.
It now says the opposite, in the operation description and again on the 409:
on `not_releasable`, READ THE ROW. The row carries the one fact that separates
the two — an assignee, or none. One extra request on an uncommon path is the
honest price of one code rather than two, and the handler's `detail` says so
too, since that is the string an operator actually reads.
THE ACCURACY ONE. CodeNotReleasable claimed "neither refusal is typed", which
is wrong and misleads the person who eventually splits it: ErrNotClaimed and
ErrNotReleasable are two distinct typed sentinels and failRelease holds both
in one case arm, so a split is a mapping change in that arm — no archaeology
and no prose-scraping. What is genuinely unavailable typed is the OBSERVATION
each refusal made, the status seen and the emptiness of the assignee, which is
why the code carries no extension member. That is a narrower claim than the
one the comment made, and it is the true one. The merge is recorded as what it
is: a deliberate start-narrow choice, kept because adding a code later is the
reversible direction.
THE WIDENING. The 200 now states it: the transition is defined over `open` AND
`in_progress`, so a row already open that STILL HOLDS AN ASSIGNEE is released
here — the state a partial write or an out-of-band status edit leaves behind,
and one a release-if-current guarding on `in_progress` did nothing for and
reported no error about. A client porting such a guard has to know that before
it ports one.
Beside it, the code-to-outcome mapping a release-if-current client wants, so
that mapping is written against a documented fact: `not_releasable`,
`precondition_failed`, `already_claimed` and `404` all mean "no release
happened, and that is not an error" — `(false, nil)` to a client spelling it
as a pair. With the reading that pair does NOT license, stated beside it,
because it is the same reading the safety finding is about: it means no
release happened, never that the claim is gone.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Review Sim <ci@beads.test>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>