beads/cmd/bd/human_proxied_server.go
Julian Knutsen 5126de8c1a
perf(bd): no git subprocesses on read-only commands (in-process repo discovery, lazy actor, backup probe) (#7400)
* perf(git): locate the repository in-process instead of running git rev-parse

Every bd invocation resolved the process-wide git context (beads directory
discovery reads it) with `git rev-parse --git-dir --git-common-dir
--show-toplevel`: one git subprocess per bd call before any work. Tools that
call bd hundreds of times (gascity's doctor issues ~250 bd calls per run)
paid it on every call.

On Linux, initGitContext now first runs discoverGitInProcess, which mirrors
Git's discovery walk (physical cwd, .git as directory or gitfile, bare-dir
check, filesystem-boundary stop) and produces byte-for-byte what rev-parse
prints for ordinary repositories (top and subdirectories), linked worktrees
(absolute or relative gitfile, symlinked targets), submodules (core.worktree
pointing back at the gitfile's directory), nested repositories and no
repository. Both answers feed one shared normalization
(gitContextFromRevParse), so downstream values are unchanged.

It declines, and git runs exactly as before, whenever it cannot prove the
answer: GIT_DIR / GIT_WORK_TREE / GIT_COMMON_DIR / GIT_CEILING_DIRECTORIES /
GIT_DISCOVERY_ACROSS_FILESYSTEM / GIT_OBJECT_DIRECTORY / GIT_CONFIG_PARAMETERS
/ GIT_CONFIG_COUNT in the environment; a symlinked .git, an invalid gitfile
or git directory, a cwd inside a git directory; repository config with
core.worktree (other than the submodule shape), core.bare, extensions,
include directives, a format version above 1 or unmodeled syntax; any path
not owned by the effective uid (safe.directory is git's to evaluate), or
root under sudo. darwin and windows keep the subprocess (case folding and
ownership semantics).

Tests compare discoverGitFrom with real `git rev-parse` output across those
layouts, pin every decline case, and check the resulting cached context
equals the subprocess-derived one.

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

* perf(cli): resolve the git user.name actor fallback only when needed

PersistentPreRunE assigned `actor = getActorWithGit()` for every command,
which, with no --actor / BEADS_ACTOR / BD_ACTOR / config actor, runs
`git config user.name`: the second git subprocess of every bd call, paid
by read-only commands that never consult the actor.

The pre-run now calls deferActorGitFallback, which marks an empty actor as
pending; currentActor (and getActorWithGit, getActor) resolve it on first
use with the unchanged priority order and store it in the global and the
CommandContext copy, as the eager assignment did. Command code reads the
actor through currentActor() instead of the raw global (a mechanical
rewrite of every read site the compiler reported; writes, the flag binding
and the pre-run config binding keep the raw variable). setActor cancels a
pending fallback. The telemetry span's bd.actor attribute is still set, but
only for a recording span, so a noop span does not force the lookup.

With the in-process repository discovery from the previous commit, a
read-only `bd list` / `bd show` now starts no git process at startup, in
the main checkout and in a linked worktree; a write asks git for user.name
once (bd create previously asked once, now still once) and records the same
value. Measured on 50 sequential `bd list` calls: 100 -> 0 git execs
(embedded and proxied-server workspaces, main checkout and worktree).

TestEmbeddedReadOnlyCommandsSpawnNoGit counts git execs through a PATH shim;
TestDeferredActorGitFallback pins the deferral semantics.

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

* fix(git): decline in-process discovery on any config line git might read differently

Review follow-ups to the in-process repository discovery.

- The repository config scanner was looser than git: a config starting with
  a UTF-8 BOM followed by `[core] worktree = /other` made git report /other
  while discovery returned the cwd repository, and `[ core ]` or `!!! = x`
  (which git refuses) were accepted. scanGitConfig (gitconfig_linux.go) now
  skips a leading BOM exactly as git does and accepts only a narrow grammar:
  blank/comment lines, `[section]` / `[section "sub"]` headers at column 0,
  and `key` / `key = value` lines inside a section, with values free of
  backslashes and unbalanced quotes. Any other line declines to git.
- Walk from getcwd(2) (syscall.Getwd), as git does, instead of os.Getwd,
  which trusts $PWD when it names the same inode (bind-mount aliases).
- Resolve relative gitfile and commondir targets without filepath.Join's
  lexical Clean, so "link/.." resolves physically like git's realpath.
- A mount-boundary stop reports its own reason in the not-a-repository
  error.
- Tests: every decline case (BOM + core.worktree, malformed headers and
  keys, keys before a section, escapes, continuations, bare repositories,
  a commondir inside .git, ...) now also checks that what startup derives
  equals git's answer; BOM and CRLF/comment configs are equivalence cases;
  the no-repository case must be answered in-process instead of skipping;
  a $PWD alias test pins the getcwd walk.
- The exec-count test drops every GIT_* variable except the config-file
  selectors, so a runner exporting GIT_CONFIG_PARAMETERS & co. cannot make
  it fail spuriously.
- Comment the two places that touch the raw actor before the deferral.

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

* perf(backup): decide embedded auto-backup's git-remote probe without git

In an embedded workspace whose backup.enabled is unset, maybeAutoBackup
asks isBackupAutoEnabled -> primeHasGitRemote after every command, reads
included. That built a RepoContext (`git -C .beads rev-parse
--git-common-dir`) and ran `git remote`: two git processes per read even
after the startup calls were gone. The throttle cannot short-circuit it:
without a remote no backup ever runs, so there is no backup state to
throttle on.

Both probes are now answered in-process when provable, and the subprocess
remains the fallback, so the decision and every warning/notice downstream
of it are unchanged:

- internal/beads getGitCommonDirForPath uses git.CommonDirInProcess (the
  discovery walk from the given directory, git's spelling).
- primeHasGitRemote uses git.HasRemoteInProcess in the same directory
  primeGitCmd would run git in (primeGitDir, factored out). It answers only
  when every config file `git remote` reads -- system (GIT_CONFIG_SYSTEM, or
  /etc/gitconfig and <prefix>/etc/gitconfig of the git on PATH), global
  (GIT_CONFIG_GLOBAL, or XDG and ~/.gitconfig) and the repository's own --
  is absent or within scanGitConfig's grammar and include-free, no
  GIT_CONFIG_PARAMETERS/COUNT/GIT_CONFIG or discovery override is set, and
  no legacy .git/remotes or .git/branches entries exist. A remote seen only
  in a system-config candidate declines. Remote names outside
  [A-Za-z0-9._-] decline.

Tests compare HasRemoteInProcess with `git remote` (no remote, local,
worktree, submodule, remote.pushDefault only, global remote, no repo) and
pin the declines (global include, malformed global, command-line config);
CommonDirInProcess is compared with `git -C dir rev-parse
--git-common-dir`. TestEmbeddedReadOnlyCommandsSpawnNoGit now runs with
backup.enabled at its default, with and without a git remote, and still
expects zero git processes for list/show.

Agent-Signature: claude-code-claude-opus-5-5-medium 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>
2026-10-08 20:08:24 +00:00

145 lines
5.9 KiB
Go

package main
import (
"context"
"errors"
"fmt"
"github.com/steveyegge/beads/internal/storage"
"github.com/steveyegge/beads/internal/storage/domain"
"github.com/steveyegge/beads/internal/storage/uow"
"github.com/steveyegge/beads/internal/types"
"github.com/steveyegge/beads/internal/ui"
"github.com/steveyegge/beads/internal/workapi"
)
// The proxied-server twins of the four `bd human` subcommands. The direct
// route reaches storage through ensureStoreActive(), which in proxied mode
// would lazily open a DIRECT store and silently bypass the proxy (bd-m7zzd) —
// so each RunE branches here BEFORE that call. Reads ride one read-only unit
// of work; each write invocation is exactly ONE RunTx with a real commit
// message, per the proxied write convention.
// proxiedHumanIssues is the proxied-server side of humanIssues: the SAME
// humanListRequest, resolved against the workspace's own status/type config
// and queried through a unit of work. It goes through openAndPrepare rather
// than hand-building a filter so `bd human list` cannot mean one thing in
// direct mode and another behind the proxy — the divergence a second
// hand-built filter here would reintroduce the next time the defaults move.
func proxiedHumanIssues(ctx context.Context, status string) ([]*types.Issue, error) {
uw, filter, err := openAndPrepare(ctx, listInput{ListRequest: humanListRequest(status)})
if err != nil {
return nil, err
}
defer uw.Close(ctx)
page, err := uw.IssueUseCase().SearchIssues(ctx, "", filter)
if err != nil {
return nil, err
}
return page.Items, nil
}
// runHumanRespondProxiedServer takes the fully-formatted comment text, not the
// raw response: `bd human respond` resolves its text sources and applies the
// "Response: " shape once, so both backends store identically-shaped comments.
func runHumanRespondProxiedServer(ctx context.Context, issueID, commentText string) error {
res, err := closeHumanProxied(ctx, issueID, commentText, "Responded", "human respond")
if err != nil {
return HandleErrorRespectJSON("%v", err)
}
if res.labelsKnown {
warnIfNotHumanLabeled(res.issue)
}
fmt.Printf("%s Bead %s closed with response.\n", ui.RenderPass("✔"), res.issue.ID)
return nil
}
// runHumanDismissProxiedServer takes the fully-formatted close reason, not the
// raw dismissal note: `bd human dismiss` resolves its text sources and applies
// the shared dismissedCloseReason prefix once for both backends.
func runHumanDismissProxiedServer(ctx context.Context, issueID, closeReason string) error {
res, err := closeHumanProxied(ctx, issueID, "", closeReason, "human dismiss")
if err != nil {
return HandleErrorRespectJSON("%v", err)
}
if res.labelsKnown {
warnIfNotHumanLabeled(res.issue)
}
fmt.Printf("%s Bead %s dismissed.\n", ui.RenderPass("✔"), res.issue.ID)
return nil
}
// humanCloseResult carries the bead as read inside the close transaction, with
// its labels when they loaded, so the caller can warn once the transaction has
// committed. labelsKnown separates "loaded, and the label is absent" from "the
// load failed" — only the first is grounds for the advisory warning.
type humanCloseResult struct {
issue *types.Issue
labelsKnown bool
}
// closeHumanProxied resolves a bead and closes it with closeReason inside ONE
// proxied-server transaction, adding comment first when it is non-empty, so
// the already-closed verdict and the close ride the same snapshot.
//
// It resolves through GetIssueOrWisp rather than proxiedRequireIssue because a
// human-labeled bead can be a WISP: `bd human list` shows the whole ephemeral
// plane, so a bead a person can see here must be one they can also answer. A
// wisp takes the wisp-side comment and close calls; the durable path is
// unchanged.
//
// Close hooks are NOT fired here. They used to need hand-wiring at each
// proxied call site, but the unit-of-work plumbing fires them itself now
// (uow.NewNotifyingProvider, bd-opisf) — buffered during the transaction and
// drained after Commit. A hand-wired call here would fire each hook twice.
func closeHumanProxied(ctx context.Context, id, comment, closeReason, commitVerb string) (humanCloseResult, error) {
if uowProvider == nil {
return humanCloseResult{}, errors.New("proxied-server UOW provider not initialized")
}
return uow.RunTxResult(ctx, uowProvider, func(ctx context.Context, uw uow.UnitOfWork) (humanCloseResult, string, error) {
src := workapi.NewUOWDetailSource(uw)
issue, isWisp, err := workapi.GetIssueOrWisp(ctx, src, id)
if errors.Is(err, storage.ErrNotFound) {
return humanCloseResult{}, "", fmt.Errorf("issue not found: %s", id)
}
if err != nil {
return humanCloseResult{}, "", fmt.Errorf("resolving issue ID %s: %w", id, err)
}
if issue.Status == types.StatusClosed {
return humanCloseResult{}, "", fmt.Errorf("issue %s is already closed", issue.ID)
}
res := humanCloseResult{issue: issue}
// Labels feed only the advisory human-label warning, so a failed load
// means no warning — not a warning that the label is missing, which
// is what ignoring the error used to produce.
if labels, lerr := src.Labels(ctx, issue.ID, isWisp); lerr == nil {
issue.Labels = labels
res.labelsKnown = true
}
if comment != "" {
var cerr error
if isWisp {
_, cerr = uw.CommentUseCase().AddCommentToWisp(ctx, issue.ID, currentActor(), comment)
} else {
_, cerr = uw.CommentUseCase().AddCommentToIssue(ctx, issue.ID, currentActor(), comment)
}
if cerr != nil {
return humanCloseResult{}, "", fmt.Errorf("adding comment: %w", cerr)
}
}
params := domain.CloseIssueParams{Reason: closeReason}
if isWisp {
_, err = uw.IssueUseCase().CloseWisp(ctx, issue.ID, params, currentActor())
} else {
_, err = uw.IssueUseCase().CloseIssue(ctx, issue.ID, params, currentActor())
}
if err != nil {
return humanCloseResult{}, "", fmt.Errorf("closing bead: %w", err)
}
return res, fmt.Sprintf("bd: %s %s", commitVerb, issue.ID), nil
})
}