Repository navigation
Description
Summary
On graft / Trail-Code 0.23.0 (macOS), the Claude Code UserPromptSubmit hook runs graft ask through execFileSync with a derived child timeout. When the working tree has drifted enough that ensureFreshGraph rebuilds inline, that rebuild is largely synchronous CPU work while holding graft/.cache/.sync.lock. The child continues running (and the hook blocks) for far longer than the intended ~4 s child budget. If the host then SIGKILLs the hook (Claude Code’s own hook timeout), the lock file is left behind naming a dead pid; later queries wait briefly and then answer from a stale graph.
This compounds the units issue in #283: a settings value of "timeout": 60 (seconds in Claude Code) is read as 60 ms of budget before overhead, then floored to 4000 ms for the child.
Environment
- Package:
@nanonets/graft/ Trail-Code 0.23.0 - OS: macOS 27.0.1 (Darwin arm64)
- Node: v26.8.1
- Hook entry:
~/.claude/helpers/graft-hooks.cjs prompt - Isolated
CLAUDE_CONFIG_DIRwith"timeout": 60on the graft UserPromptSubmit command hook
Observed (minimal repro)
- Cold
graft buildof ~3500 Python files: ~28 s, peak RSS ~1.1 GiB. promptAskTimeout(dir)with installed timeout60→ 4000.- Wipe extract cache + modify 500 files; run the prompt hook as Claude Code does.
- Hook wall: ~43 s (expected child budget 4 s).
graft askchild remained in runnable state and held.sync.lockuntil ~39 s.- Hook produced no
additionalContext(child timed out / null JSON).
- Same setup under
gtimeout -s KILL 10:.sync.locksurvived with a dead pid.- Next
graft askprinted:a graph rebuild is already in flight — answering from the current graph(~3.8 s), lock still present.
- Control:
GRAFT_NO_REFRESH=1on the same hook → ~2.3 s with a normal retrieval pack.
Repro script (synthetic 3500-file repo, isolated CLAUDE_CONFIG_DIR; does not touch ~/.claude)
# 0) Preconditions
node -e "console.log(require(require('os').homedir()+'/.local/lib/node_modules/@nanonets/graft/package.json').version)"
# expect 0.23.0
uptime # wait until 1-min load <= 20
BASE=$(mktemp -d /tmp/graft-repro-XXXX)
R=$BASE/repo
CFG=$BASE/claude-config
mkdir -p "$R" "$CFG" "$BASE/trail-home"
git -C "$R" init -q
# 1) Synthetic tree (~3500 Python files; adjust N if build time not 20–60s)
N=3500
i=0
while [ "$i" -lt "$N" ]; do
d=$((i / 100)); mkdir -p "$R/src/mod_$d"
cat > "$R/src/mod_$d/file_$i.py" <<EOF
class ParseError(Exception): pass
def tokenize(s): return s.split()
def parse_expr(tokens, pos=0):
if pos >= len(tokens): raise ParseError("eof")
return tokens[pos], pos+1
def handle_errors(src):
try: return parse_expr(tokenize(src))[0]
except ParseError as e: return {"error": str(e), "mod": $i}
EOF
i=$((i + 1))
done
git -C "$R" add -A
git -C "$R" -c user.email=repro@local -c user.name=repro commit -q -m 'repro tree'
# 2) Cold build + RSS
cd "$R"
/usr/bin/time -l graft build .
# record real time + "maximum resident set size"
# 3) Confirm timeout math (isolated settings; does not touch ~/.claude)
cat > "$CFG/settings.json" <<'EOF'
{
"hooks": {
"UserPromptSubmit": [{
"hooks": [{
"type": "command",
"command": "node ~/.claude/helpers/graft-hooks.cjs prompt",
"timeout": 60
}]
}]
}
}
EOF
CLAUDE_CONFIG_DIR=$CFG node --input-type=module -e "
import { promptAskTimeout } from 'file://'+process.env.HOME+'/.local/lib/node_modules/@nanonets/graft/dist/claude/hooks.js';
console.log('promptAskTimeoutMs', promptAskTimeout(process.argv[1]));
" "$R"
# expect 4000
# 4) Force long inline rebuild (drift + wipe extract cache)
rm -f "$R"/graft/.cache/extract.*.json
i=0; while [ "$i" -lt 500 ]; do
f="$R/src/mod_$((i/100))/file_$i.py"
[ -f "$f" ] && printf '\n# drift %s\n' "$i" >> "$f"
i=$((i+1))
done
# 5) Run hook exactly as Claude Code (isolated config)
export CLAUDE_CONFIG_DIR=$CFG CLAUDE_PROJECT_DIR=$R TRAIL_HOME=$BASE/trail-home
# In another terminal while this runs: ps -ax -o pid,ppid,etime,state,command | grep 'cli.js ask'
printf '%s' '{"prompt":"how does the parser handle errors","session_id":"repro","hook_event_name":"UserPromptSubmit"}' \
| /usr/bin/time -p node ~/.claude/helpers/graft-hooks.cjs prompt
# expect wall >> 4s; often empty stdout
# 6) SIGKILL simulation + stale lock
rm -f "$R"/graft/.cache/extract.*.json
i=0; while [ "$i" -lt 500 ]; do
f="$R/src/mod_$((i/100))/file_$i.py"
[ -f "$f" ] && printf '\n# drift2 %s\n' "$i" >> "$f"
i=$((i+1))
done
printf '%s' '{"prompt":"how does the parser handle errors","session_id":"repro-kill","hook_event_name":"UserPromptSubmit"}' \
| gtimeout -s KILL 10 /usr/bin/time -p node ~/.claude/helpers/graft-hooks.cjs prompt || true
cat "$R/graft/.cache/.sync.lock"
# expect JSON with dead pid; kill -0 <pid> fails
/usr/bin/time -p graft ask "how does the parser handle errors" "$R" --json -n 1
# stderr note: rebuild already in flight; answers from current graph
# 7) Control
rm -f "$R/graft/.cache/.sync.lock" # optional cleanup before control
export GRAFT_NO_REFRESH=1
printf '%s' '{"prompt":"how does the parser handle errors","session_id":"repro-control","hook_event_name":"UserPromptSubmit"}' \
| /usr/bin/time -p node ~/.claude/helpers/graft-hooks.cjs prompt
# expect ~1–3s and non-empty hookSpecificOutput
# 8) Cleanup
rm -rf "$BASE"Likely mechanism
dist/claude/hooks.js:graftJson→execFileSync(..., { timeout: promptAskTimeout(dir) }).promptAskTimeoutsubtracts 2000 from the installed number and floors at 4000, without converting Claude Code seconds → ms (settings-merge: hook timeouts written in ms (Claude Code reads seconds) and footerLinksRegexes written to project settings where it is ignored #283).dist/graph/refresh.js:ensureFreshGraphtakes.sync.lock, installsreleaseOnSignal, thenawait buildGraph(...). Synchronous parse/build work can delay handling of SIGTERM until the rebuild finishes, so the child (and parent hook) overrun the timeout.releaseOnSignalnever runs if the process is SIGKILL’d.dist/util/state.js:acquireLockIntreats an existing lock as free only afterLOCK_STALE_MS(300000 ms); it does not check whetherpidis alive.
Expected
- Child timeout should track Claude Code’s seconds budget (or settings should store ms consistently) so a 60 s hook allows a proportionate ask/rebuild window — or rebuild should not run synchronously under the prompt hook’s short child budget.
- SIGTERM during rebuild should abort promptly and always release
.sync.lock. - A lock whose pid is dead should be reclaimable immediately (or within seconds), not up to five minutes of “rebuild in flight” stale answers.
Related
- settings-merge: hook timeouts written in ms (Claude Code reads seconds) and footerLinksRegexes written to project settings where it is ignored #283 — hook timeouts written in ms vs Claude Code seconds
- post-edit hook: a fixed 8s
graft checkchild inside a 10s budget, and mergeGraftSettings drops a hand-raised timeout on every wiring refresh #366 — the same fixed-child-budget pattern in the post-edit hook - Refresh the graph before every query, incrementally #25 — refresh the graph before every query (introduced the inline rebuild on the prompt path)
Workaround
GRAFT_NO_REFRESH=1 in front of the prompt hook command (it then answers from the current graph; the Stop hook's background sync still refreshes it). graft init rewrites the hook command, so this has to be re-applied after every init.
Activity
hr1juldey commentedon Nov 18, 2026
Picked up the stale-lock half of this. On current main the lock's pid-liveness check only ran after the 5-minute age floor, so the SIGKILLed hook's fresh dead-pid lock kept every query waiting and answering from the stale graph for up to five minutes — the repro's step 4. Draft: #861 makes a dead owner lose the lock immediately while keeping the age rules for live owners (and
lockHeldByOthermoves with it, so the next caller rebuilds instead of waiting). The child-budget half (timeout units, compounding #283) is a separate surface and deliberately untouched.fix(lock): a dead pid loses its lock immediately, not after five minu…
fix(lock): a dead pid loses its lock immediately, not after five minu…