[johnybradshaw]

Summary

On mid-size repos, the post-edit hook blocks every Write/Edit for about 8 seconds and then reports 0 stale (#483). This PR moves the stale count off the edit path. It also fixes a second bug: graft check --json truncates its own output on a pipe.

1. fix(claude): count stale nodes in a detached child (#483)

handlePostEdit ran graft check synchronously under the 8s CHILD_TIMEOUT_MS. On a ~900-file repo the check alone takes 6.5–9.5s, CPU-bound. Across 127 recorded post-edit runs, each took 6.3–8.4s, and a timeout was written as staleCount: 0.

Now handlePostEdit sets dirty and lastFile and spawns dist/claude/stale-count-run.js detached, the way handleStop spawns sync-run.js. It uses a GRAFT_TEST_STALE_COUNT_RUN seam, like GRAFT_TEST_SYNC_RUN. The runner works as follows:

• It checks under a 30s budget, not sync-run's 120s. A tree too large to check in 30s would otherwise burn a core for two minutes per edit burst and still get nothing.
• It takes its own lock, .cache/stale-check.lock, so one check runs at a time and the sync lock is never held. A lock older than the budget plus 10s is reclaimed, and the runner only removes a lock whose pid is its own.
• It never invents a 0. On a timeout or unparseable output it leaves the last count standing.
• It never overwrites a newer answer. It skips the write if the graph is clean, a sync is running, or syncedAt is later than the check's start.
• The blast-radius, overlap and learnings output is unchanged.

2. fix(cli): let check --json finish writing before a stale exit

I found this while verifying the change end to end. Even when the check finished inside the budget, the count stayed at 0. graft check called process.exit(1) straight after console.log. On POSIX, stdout to a pipe is asynchronous, so whatever was still queued was dropped:

reader	bytes received (full output 530,396)
file redirect	530,396
| wc -c	65,536
execFileSync / spawnSync	146,176

graph.pendingIds lists every unsummarised node, so any repo of a few thousand symbols built without --deep produces output past that size. post-edit's parse then failed and wrote 0. Anyone piping graft check --json into jq hits the same thing.

The fix is process.exitCode = 1, which other commands in cli.ts already use. One side effect to be aware of: postAction now runs for a failing check, so the query telemetry event counts stale checks as well as clean ones. Before this, only passing checks were counted.

Relation to #485 and #450

• fix: hook timeout units are seconds not ms (#283, #480); post-edit check budget and retry-storm (#366, #483) #485 handles post-edit hook burns its full 8 s child timeout on every edit in large repos and discards the result #483 by skipping the synchronous check after it has timed out once (checkTimedOut). On a repo whose check takes around 8s, that still blocks the first edit after every sync, and the count is still subject to the truncation above. This PR replaces that part of fix: hook timeout units are seconds not ms (#283, #480); post-edit check budget and retry-storm (#366, #483) #485; the timeout-units fixes there (settings-merge: hook timeouts written in ms (Claude Code reads seconds) and footerLinksRegexes written to project settings where it is ignored #283/Claude Code hook timeouts written in milliseconds; Claude Code reads seconds #480) are independent. Both PRs touch handlePostEdit, so whichever lands second needs a small rebase. I'm happy to do it.
• fix(claude): skip post-edit's graft check when no graft statusline reads it #450 only skips the check when no graft statusline reads it. When the statusline is wired, which is the case that needs the count, it doesn't help.

Test plan

• New claude-hooks tests. post-edit spawns the runner with the repo dir and returns without waiting on it, leaving staleCount for the child. runStaleCount covers: recovering JSON from a non-zero exit; releasing its lock; keeping the last count on null; each of the three sync guards, plus a sync from before the check not blocking; the live lock; and reclaiming a stale lock. The existing GRAFT_DIR --dir test moved to the runner.
• Mutation-checked. Removing each guard, the lock check or release, the stale-lock reclaim, stdout recovery, --dir, or the spawn, or making the spawn synchronous, fails at least one test.
• New cli-check-pipe test. A 12,000-symbol repo is made stale, graft check --json is read through spawnSync, and the test asserts exit status 1 and that the JSON parses. Without the fix it fails with SyntaxError: Unterminated string in JSON at position 146176. A size assertion makes the fixture fail loudly if it ever shrinks below what a pipe can hide.
• Full suite: 2530 pass. All failures are also on main in the same environment (unvendored grammars, symlink and git-tracking cases), apart from wall-clock-bound tests that failed under full-suite load and pass alone, before and after: trace-shell, skills-answers-prompt, upkeep-hooks.
• End to end, with the built dist, on a copy of a real 553-file repo after an edit: the hook took 220–250ms, against 8.27s for unpatched 0.23.1 on a similar repo. The background check finished in about 11s, released its lock, and wrote staleCount: 1, matching graft check.

Fixes #483

🤖 Generated with Claude Code

[trailhq-graft]

🌱 graft blast radius

2 areas changed → 3 areas can be affected. 15 dependent symbols, depth 2.
Tests: no test reaches Check Command; 1 area updated its tests.
Tag: @anirudhkumar-nanonets — Hook I/O, Check Command · @shhdwi — 5 of 5 areas

flowchart TB
  A0(("Skill Lifecycle Hooks<br/>8 symbols"))
  A1(("Lesson Gating<br/>6 symbols"))
  A2(("Skill Session Tracing<br/>1 symbol"))
  classDef reached fill:#D9EDF3,stroke:#3AA7C9,stroke-width:1.5px,color:#0E313C;
  class A0,A1,A2 reached;

Loading

Can be affected	Symbols	Nearest hop	Reached from
Skill Lifecycle Hooks	8	src/claude/prompt-hook.ts:L1-L298 prompt-hook.ts — imports, depth 1	Hook I/O
Lesson Gating	6	src/skills/lesson-gate.ts:L1-L320 lesson-gate.ts — imports, depth 2	Hook I/O
Skill Session Tracing	1	src/trace/skill-session.ts:L1-L265 skill-session.ts — imports, depth 1	Hook I/O

Who knows this code — 2 people across 5 areas

Area	Who knows it
Hook I/O · changed	@anirudhkumar-nanonets — 16 commits, last yesterday · @shhdwi — 13 commits, last today
Check Command · changed	@anirudhkumar-nanonets — 61 commits, last yesterday · @shhdwi — 39 commits, last today
Skill Lifecycle Hooks · affected	@shhdwi — 2 commits, last today
Lesson Gating · affected	@shhdwi — 2 commits, last today
Skill Session Tracing · affected	@shhdwi — 1 commit, last yesterday

Ownership is git history over each area's own files, weighted towards recent work (120-day half-life). Merge commits and bots are dropped, and you are dropped from your own PR. A name with no @ has no GitHub handle in its commit email — tag them by hand, or add a .mailmap entry. A suggestion from history, not a CODEOWNERS rule.

All 15 dependent symbols, grouped by area

Skill Lifecycle Hooks — 8 symbols in 8 files

• src/claude/prompt-hook.ts:L1-L298 — prompt-hook.ts (imports, depth 1)
  7: * out of time must not take the learnings down with it. The graph ask (`graft
• src/claude/prompt-skills.ts:L1-L327 — prompt-skills.ts (imports, depth 1)
  5: * 3-character floor and before `graft ask`.
• src/claude/stop.ts:L1-L213 — stop.ts (imports, depth 1)
  5: * reply said what graft saved, and one background graph sync when an edit
• src/claude/post-bash.ts:L1-L186 — post-bash.ts (imports, depth 2)
• src/claude/skills-hook-util.ts:L1-L83 — skills-hook-util.ts (imports, depth 2)
• src/claude/start-skills.ts:L1-L269 — start-skills.ts (imports, depth 2)
• src/claude/stop-skills.ts:L1-L236 — stop-skills.ts (imports, depth 2)
• src/claude/walls-hook.ts:L1-L247 — walls-hook.ts (imports, depth 2)

Lesson Gating — 6 symbols in 6 files

• src/skills/lesson-gate.ts:L1-L320 — lesson-gate.ts (imports, depth 2)
  63: subagent: `${NOT}only the main session can keep a lesson.`,
• src/skills/offer-lesson.ts:L1-L125 — offer-lesson.ts (imports, depth 2)
• src/skills/offer-moment.ts:L1-L307 — offer-moment.ts (imports, depth 2)
• src/skills/save-gate.ts:L1-L363 — save-gate.ts (imports, depth 2)
  159: * The main session's tool calls in the transcript's last 4 MB, oldest first, as the host's dialect reads them (4.16):
• src/skills/save-verify.ts:L1-L390 — save-verify.ts (imports, depth 2)
  63: subagent: `${NOT}only the main session can keep a skill.`,
• src/skills/update-gate.ts:L1-L345 — update-gate.ts (imports, depth 2)
  65: subagent: `${NOT}only the main session can change a skill.`,

Skill Session Tracing — 1 symbol in 1 file

• src/trace/skill-session.ts:L1-L265 — skill-session.ts (imports, depth 1)
  21: import { payloadSessionId } from '../claude/hook-io.js';

Test signal per changed area — 1 ✓ · 1 ✗

Reached = a node under a test path has a resolved edge into the changed symbol. It undercounts anything called indirectly — through a CLI, a spawned process or a dynamic import — so read a low ratio as “look here”, never as a coverage gate.

• ✓ Hook I/O — 1 of 7 reached · 1 test file changed here: test/claude-hooks.test.ts
  • not reached: staleCountInBackground, handlePostEdit, acquire, countStale, main, release
• ✗ Check Command — 0 of 1 reached · no test file reaches it
  • not reached: runCheckCommand

52 test suites also reference this code

52 symbols, kept out of the diagram and the table so they cannot crowd out the areas a reviewer has to look at.

• test/claude-copy-dashes.test.ts
• test/claude-hook-timeouts.test.ts
• test/claude-hooks-perf.test.ts
• test/claude-hooks-skills.test.ts
• test/claude-host-standdown.test.ts
• test/claude-pre-bash-skills.test.ts
• test/claude-prompt-learnings.test.ts
• test/claude-prompt-private.test.ts
• test/claude-session-start-learnings.test.ts
• test/claude-skills-journey-fixtures.ts
• test/claude-stop-moments.test.ts
• test/claude-stop-trace.test.ts
• test/claude-stop.test.ts
• test/claude-tally.test.ts
• test/hosts-codex-hooks.test.ts
• test/skills-agent-trust.test.ts
• test/skills-anchors-wide-review.test.ts
• test/skills-bound-values.test.ts
• test/skills-check-readonly.test.ts
• test/skills-consent-fixtures.ts
• …32 more

_{graft blast · origin/main...HEAD · depth 2 · 6 changed files}

Open the interactive graph → — click an area to see its dependent symbols at file:line.