fix(#2839): transactional cleanup tail for /gsd-code-review-fix (#2846)

* fix(#2839): make /gsd-code-review-fix cleanup transactional

Cleanup tail in agents/gsd-code-fixer.md previously did 'git worktree
remove' without any recovery marker. If the process was killed between
fix commits and worktree removal, the orphan worktree + branch survived
with no resume path — the next run had no way to discover or finish
the cleanup.

Introduce a recovery sentinel at ${phase_dir}/.review-fix-recovery-pending.json
with strict ordering:
- Sentinel written AFTER 'git worktree add' succeeds (never points at a
  worktree that does not exist).
- Sentinel removed ONLY AFTER 'git worktree remove' returns successfully
  (interruption between commits and removal leaves a sentinel behind).
- New runs detect a pre-existing sentinel, force-remove the recorded
  orphan worktree, then drop the stale sentinel before continuing —
  making the agent self-healing after a crash.

Closes #2839

* fix(#2839): harden sentinel JSON parse and scope ordering assertion

Address CodeRabbit review feedback on PR #2846:

- agents/gsd-code-fixer.md: Guard the recovery-sentinel JSON parse with
  try/catch so a corrupted/truncated sentinel (a realistic crash artifact)
  emits a warning and yields an empty prior_wt instead of aborting setup.
  This preserves the self-healing recovery path even when the sentinel
  itself is the casualty of the original crash.

- tests/bug-2839-review-fix-transactional-cleanup.test.cjs: Scope the
  cleanup-ordering assertion to the cleanup-tail section of the
  setup_worktree step rather than first global occurrences. Previously
  the assertion could pass on pre-recovery references even if cleanup-tail
  ordering regressed. The regex also now accepts the shell-variable form
  (\`rm -f \"\$sentinel\"\`) used in the cleanup tail.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
Tom Boucher
2026-04-29 07:56:32 -04:00
committed by GitHub
parent c3a42d66f9
commit 7fae804296
2 changed files with 247 additions and 7 deletions

View File

@@ -214,32 +214,89 @@ If a finding references multiple files (in Fix section or Issue section):
This agent runs as a background process that makes commits. Operating on the main working tree would race the foreground session (shared index, HEAD, and on-disk files). Instead, every instance runs in its own isolated worktree.
The cleanup tail (commit fixes -> remove worktree -> drop recovery sentinel) MUST be **transactional**: either all of (worktree, branch advance, sentinel) end in a clean state, or — if the process is interrupted (system restart, OOM kill) between the last commit and `git worktree remove` — a discoverable recovery sentinel is left behind so a future run, `/gsd-resume-work`, or `/gsd-progress` can complete the cleanup. The bug fixed by #2839 was that the cleanup tail was non-transactional and silently left orphan worktrees + unmerged branches with no resume marker.
```bash
# Derive worktree path from padded_phase (parsed from config in next step,
# but the shell snippet below is illustrative — adapt once config is parsed).
# In practice: parse padded_phase from config first, then run:
branch=$(git branch --show-current)
test -n "$branch" || { echo "Detached HEAD is not supported for review-fix (#2686)"; exit 1; }
# Recovery-sentinel handling (#2839):
# Path is ${phase_dir}/.review-fix-recovery-pending.json. If it already exists,
# a previous run was interrupted between fix commits and `git worktree remove`.
# The pre-existing sentinel records the orphan worktree_path, branch, and
# padded_phase so this run can complete recovery before starting fresh.
sentinel="${phase_dir}/.review-fix-recovery-pending.json"
if [ -f "$sentinel" ]; then
echo "Detected pre-existing recovery sentinel from a prior interrupted run: $sentinel"
prior_wt=$(node -e '
const fs = require("fs");
try {
const parsed = JSON.parse(fs.readFileSync(process.argv[1], "utf-8"));
process.stdout.write(parsed.worktree_path || "");
} catch (err) {
process.stderr.write(`Warning: malformed recovery sentinel ${process.argv[1]}: ${err.message}\n`);
process.stdout.write("");
}
' "$sentinel")
if [ -n "$prior_wt" ] && git worktree list --porcelain | grep -q "^worktree $prior_wt$"; then
echo "Removing orphan worktree from prior run: $prior_wt"
git worktree remove "$prior_wt" --force || true
fi
rm -f "$sentinel"
fi
wt=$(mktemp -d "/tmp/sv-${padded_phase}-reviewfix-XXXXXX")
git worktree add "$wt" "$branch"
# Write the recovery sentinel ONLY AFTER `git worktree add` succeeds.
# Writing it before would leave a sentinel pointing at a worktree that does
# not exist if `git worktree add` itself failed.
node -e '
const fs = require("fs");
const [sentinelPath, worktree_path, branch, padded_phase] = process.argv.slice(1);
fs.writeFileSync(sentinelPath, JSON.stringify({
worktree_path,
branch,
padded_phase,
started_at: new Date().toISOString()
}, null, 2));
' "$sentinel" "$wt" "$branch" "$padded_phase"
cd "$wt"
```
Concrete steps:
1. Parse `padded_phase` from the `<config>` block (needed for the path).
1. Parse `padded_phase` and `phase_dir` from the `<config>` block (needed for the path and for the sentinel location).
2. Resolve the current branch: `branch=$(git branch --show-current)`. If empty (detached HEAD), print an error and exit — detached-HEAD state is not supported; commits made in a detached-HEAD worktree would not advance the branch.
3. Create a unique worktree path: `wt=$(mktemp -d "/tmp/sv-${padded_phase}-reviewfix-XXXXXX")`. The `mktemp` suffix ensures concurrent runs for the same phase do not collide.
4. Run `git worktree add "$wt" "$branch"` — this attaches the worktree to the current branch so commits advance it.
5. All subsequent file reads, edits, and commits happen inside `$wt`.
3. **Recovery check (#2839):** If `${phase_dir}/.review-fix-recovery-pending.json` already exists, a prior run was interrupted. Parse the JSON, attempt to remove the orphan worktree it points at (best-effort, with `--force`), then delete the stale sentinel before continuing. This makes a re-run of `/gsd-code-review-fix` self-healing.
4. Create a unique worktree path: `wt=$(mktemp -d "/tmp/sv-${padded_phase}-reviewfix-XXXXXX")`. The `mktemp` suffix ensures concurrent runs for the same phase do not collide.
5. Run `git worktree add "$wt" "$branch"` — this attaches the worktree to the current branch so commits advance it.
6. **Write the recovery sentinel** at `${phase_dir}/.review-fix-recovery-pending.json` containing `{worktree_path, branch, padded_phase, started_at}`. Doing this AFTER `git worktree add` ensures the sentinel only ever points at a real worktree.
7. All subsequent file reads, edits, and commits happen inside `$wt`.
**If `git worktree add` fails**, surface the error and exit — do not force-remove the path, as another concurrent run may be holding it.
**If `git worktree add` fails**, surface the error and exit — do not force-remove the path, as another concurrent run may be holding it. Do not write the sentinel (the worktree does not exist).
**Cleanup tail (transactional, ALWAYS — even on failure):** After writing REVIEW-FIX.md and before returning to the orchestrator, run the two-step cleanup in this exact order:
**Cleanup (ALWAYS — even on failure):** After writing REVIEW-FIX.md and before returning to the orchestrator, run:
```bash
# Step 1: drop the worktree FIRST. If this succeeds and the process is then
# killed, the next run finds a sentinel pointing at a worktree that no longer
# exists — the recovery branch handles this gracefully (best-effort remove +
# sentinel delete). If we reversed the order (sentinel removed first, then
# worktree remove), an interruption between the two steps would leave NO
# sentinel and an orphan worktree — exactly the bug from #2839.
git worktree remove "$wt" --force
# Step 2: drop the recovery sentinel ONLY after `git worktree remove` returns
# successfully. This atomic-ish ordering is what makes the cleanup tail
# transactional from the orchestrator's perspective.
rm -f "$sentinel"
```
This cleanup is unconditional — register it mentally as a finally-block obligation. If the agent exits early (config error, no findings, etc.), still run `git worktree remove "$wt" --force` before exit.
This cleanup is unconditional — register it mentally as a finally-block obligation. If the agent exits early (config error, no findings, etc.), still run the two-step cleanup tail (`git worktree remove "$wt" --force` followed by `rm -f "$sentinel"`) before exit. The sentinel must NEVER be removed before `git worktree remove` succeeds.
</step>
<step name="load_context">
@@ -473,6 +530,8 @@ _Iteration: {N}_
**ALWAYS run inside the isolated worktree** — set up via `branch=$(git branch --show-current)` + `wt=$(mktemp -d "/tmp/sv-${padded_phase}-reviewfix-XXXXXX")` + `git worktree add "$wt" "$branch"` at the very start (see `setup_worktree` step). Using `mktemp` ensures concurrent runs do not collide. Attaching to `$branch` (not `HEAD`) ensures commits advance the branch. Every file read, edit, and commit must happen inside `$wt`. Run `git worktree remove "$wt" --force` unconditionally when done (treat it as a finally block). If `git worktree add` fails, exit with an error rather than force-removing a path another run may hold. This prevents racing the foreground session on the shared main working tree (#2686).
**ALWAYS run the transactional cleanup tail in order** (#2839): `git worktree remove "$wt" --force` MUST happen BEFORE `rm -f "$sentinel"` (the recovery sentinel at `${phase_dir}/.review-fix-recovery-pending.json`). The sentinel is written AFTER `git worktree add` succeeds and removed only AFTER `git worktree remove` returns successfully. This ordering is what makes the cleanup tail transactional — an interruption between commits and `git worktree remove` leaves the sentinel behind so a future run, `/gsd-resume-work`, or `/gsd-progress` can detect and complete the recovery. Reversing the order recreates the orphan-worktree bug.
**ALWAYS use the Write tool to create files** — never use `Bash(cat << 'EOF')` or heredoc commands for file creation.
**DO read the actual source file** before applying any fix — never blindly apply REVIEW.md suggestions without understanding current code state.

View File

@@ -0,0 +1,181 @@
/**
* Regression test for bug #2839
*
* /gsd-code-review-fix cleanup tail is non-transactional. If the agent is
* interrupted (system restart, OOM kill) AFTER the last fix commit but
* BEFORE `git worktree remove`, the worktree is orphaned in
* `git worktree list`, the agent's branch is left with unmerged commits,
* and STATE.md is never advanced. To anyone reading main only, the phase
* looks "ready to plan" while critical fixes sit on a dangling branch.
*
* Fix: introduce a recovery sentinel JSON at
* ${PHASE_DIR}/.review-fix-recovery-pending.json
* The sentinel is written AFTER `git worktree add` succeeds and
* REMOVED only after `git worktree remove` completes, so the cleanup
* tail is transactional from the orchestrator's perspective. If the
* process dies in between, the sentinel is left behind pointing at the
* orphan worktree and branch — a future run, /gsd-resume-work, or
* /gsd-progress can detect and complete the recovery.
*/
'use strict';
// allow-test-rule: source-text-is-the-product
// The gsd-code-fixer agent's working instructions ARE the product — Claude
// follows them at runtime. Structural assertions over the markdown source
// test the deployed contract. See bug-2686 for the same pattern.
const { describe, test, before } = require('node:test');
const assert = require('node:assert/strict');
const fs = require('fs');
const path = require('path');
const SENTINEL_NAME = '.review-fix-recovery-pending.json';
function parseFrontmatter(content) {
const match = content.match(/^---\n([\s\S]*?)\n---/);
if (!match) return null;
const body = match[1];
const out = {};
for (const line of body.split('\n')) {
const m = line.match(/^([a-zA-Z_]+):\s*(.*)$/);
if (m) out[m[1]] = m[2].trim();
}
return out;
}
function extractStep(content, stepName) {
const re = new RegExp(`<step\\s+name="${stepName}">([\\s\\S]*?)</step>`);
const m = content.match(re);
return m ? m[1] : null;
}
describe('bug-2839: /gsd-code-review-fix cleanup is transactional', () => {
let agentPath;
let agentContent;
let frontmatter;
before(() => {
agentPath = path.join(__dirname, '..', 'agents', 'gsd-code-fixer.md');
assert.ok(fs.existsSync(agentPath), 'agents/gsd-code-fixer.md must exist');
agentContent = fs.readFileSync(agentPath, 'utf-8');
frontmatter = parseFrontmatter(agentContent);
assert.ok(frontmatter, 'agent must have YAML frontmatter');
});
test('agent declares a recovery sentinel filename', () => {
assert.ok(
agentContent.includes(SENTINEL_NAME),
`gsd-code-fixer.md must reference the recovery sentinel ${SENTINEL_NAME} so an interrupted cleanup tail is discoverable (#2839)`
);
});
test('sentinel is written inside setup_worktree, after git worktree add', () => {
const setupStep = extractStep(agentContent, 'setup_worktree');
assert.ok(setupStep, 'setup_worktree step must exist');
assert.ok(
setupStep.includes(SENTINEL_NAME),
`setup_worktree must reference ${SENTINEL_NAME} so the sentinel is created at the start of the run (#2839)`
);
const addPos = setupStep.indexOf('git worktree add');
assert.ok(addPos !== -1, 'setup_worktree must contain `git worktree add`');
// The sentinel WRITE (not just a reference) must come after `git worktree add`.
// Earlier references are allowed (e.g. recovery check for a stale sentinel
// from a prior interrupted run). Look for an explicit write — either a
// shell `>`/`>>` redirection, a `node -e` invocation that uses
// `fs.writeFileSync(...sentinel...)`, or a `Write` tool reference.
const writeIdx = (() => {
const candidates = [
/fs\.writeFileSync\([^)]*sentinel/,
/>\s*"?\$sentinel/,
/>\s*"?\$\{sentinel\}/,
/Write the recovery sentinel/i,
];
let earliest = -1;
for (const re of candidates) {
const m = re.exec(setupStep);
if (m && (earliest === -1 || m.index < earliest)) earliest = m.index;
}
return earliest;
})();
assert.ok(
writeIdx !== -1,
'setup_worktree must explicitly describe writing the sentinel (#2839)'
);
assert.ok(
addPos < writeIdx,
'sentinel must be written AFTER `git worktree add` succeeds (#2839)'
);
});
test('sentinel records worktree path, branch, and padded_phase as JSON fields', () => {
for (const key of ['worktree_path', 'branch', 'padded_phase']) {
assert.ok(
agentContent.includes(key),
`recovery sentinel must record \`${key}\` so a future /gsd-resume-work or /gsd-progress can locate the orphan state (#2839)`
);
}
});
test('sentinel removal happens only AFTER git worktree remove succeeds', () => {
const setupStep = extractStep(agentContent, 'setup_worktree');
assert.ok(setupStep, 'setup_worktree step must exist');
const cleanupAnchor = setupStep.lastIndexOf('Cleanup tail (transactional');
assert.ok(cleanupAnchor !== -1, 'setup_worktree must document cleanup-tail section');
const cleanupSection = setupStep.slice(cleanupAnchor);
const removeIdx = cleanupSection.indexOf('git worktree remove "$wt" --force');
assert.ok(removeIdx !== -1, 'cleanup-tail must remove worktree');
// Within the cleanup-tail section, accept either a literal-filename form
// (`rm -f .../.review-fix-recovery-pending.json`) or a shell-variable form
// referring to the previously-declared `sentinel` variable
// (`rm -f "$sentinel"` / `rm -f "${sentinel}"`).
const escapedName = SENTINEL_NAME.replace(/\./g, '\\.');
const sentinelRemovalRe = new RegExp(
`(rm\\s+(?:-f\\s+)?[^\\n]*(?:${escapedName}|\\$\\{?sentinel\\}?)|unlink[^\\n]*(?:${escapedName}|\\$\\{?sentinel\\}?))`
);
const sentinelRemovalMatch = sentinelRemovalRe.exec(cleanupSection);
assert.ok(
sentinelRemovalMatch,
`agent must remove the sentinel file (rm or unlink ${SENTINEL_NAME}) as part of the cleanup tail (#2839)`
);
const sentinelRemovalIdx = sentinelRemovalMatch.index;
assert.ok(
removeIdx < sentinelRemovalIdx,
'cleanup ordering must be: `git worktree remove` BEFORE sentinel removal (#2839)'
);
});
test('agent documents detection of pre-existing sentinel from a prior interrupted run', () => {
const lower = agentContent.toLowerCase();
const mentionsRecovery =
lower.includes('stale sentinel') ||
lower.includes('existing sentinel') ||
lower.includes('previous sentinel') ||
lower.includes('prior run') ||
lower.includes('pre-existing sentinel') ||
lower.includes('recovery');
assert.ok(
mentionsRecovery,
'agent must describe how it handles a pre-existing sentinel from a previous interrupted run (#2839)'
);
});
test('cleanup-tail obligation is documented as transactional / atomic', () => {
const lower = agentContent.toLowerCase();
const mentionsTransactional =
lower.includes('transactional') ||
lower.includes('atomic cleanup') ||
lower.includes('cleanup tail');
assert.ok(
mentionsTransactional,
'agent must document the cleanup tail as transactional/atomic (#2839)'
);
});
});