daily-fix: Step 10d lint-gate pending-call wedge

Workflow gap

Across 2026-08-05 → 08-06, at least 12 autonomous /issue sessions wedged at the Step 10d pre-push lint-gate Bash dispatch: the orchestrator issued the gate command (multi-KB compound beginning sudo -n choom -n -600 -p $$ ...) and NO tool_result ever arrived — not even the instant "Async task launched" ack on the run_in_background: true calls. Each session sat ALIVE-BUT-STALLED 1.2–2.4 h until the watcher's stall fence respawned it (the pending call is recorded as "The user doesn't want to proceed with this tool use" at session stop); several respawns re-wedged at the same step.

Affected (transcript-verified pending→rejected pairs, three independent miners): #1992 (×3 incl. 12:29→13:53Z), #1996 (17:03→19:15Z), #2004 (06:41→07:53Z, 08:52→10:13Z), #2006 (19:49→21:14Z, 23:10→00:33Z), #2022 (06:40→08:33Z, 10:34→11:53Z — the second ended by a manual rejection), #2079 (07:40→10:03Z), #2085 (14:20→16:43Z, 17:03→19:23Z, 22:19→00:43Z), #2103 (00:52→03:13Z, 03:44→05:53Z), #2104 (03:52→06:13Z 08-06). Both foreground AND background dispatch shapes wedged; the same command shape also ran clean in sibling sessions the same day (e.g. #1996 at 19:48Z bg, exit 0, choom=ok) — intermittent. Aggregate: roughly 15–20 h of blocked completion wall-time, multiple watcher respawn cycles (#1992: 3; #2004: 7 per the fleet-ops session), one manual Thomas intervention, and the falsified #2084 filing ("Bound the session-auto-respawn lane") was this wedge's downstream symptom.

verified-at-filing (all run 2026-08-06T07:1xZ on main):

• uv run python -c "json.load(open('.claude/settings.json'))" hook read → 14 PreToolUse hook entries (9 matching Bash), NONE carries a timeout field.
• grep -n 'choom\|sudo' .claude/settings.json → no allowlist entry for the sudo -n choom prelude.
• Open related tasks: #1994 (blocked — "phase-done lint test hangs at interpreter exit"), #2039 (proposed — "inline lint gate defers under high VM load"), #2084 (proposed, premise falsified — respawn-loop symptom of this wedge). None covers the pending-call wedge itself.

Additional miner-6 evidence (raises the count to ~18 events): six MORE sessions (#2022 08:51→10:13Z, #1992 08:13→10:33Z, #2085 19:41→22:03Z, #2087 19:47→22:13Z, #2006 21:34→22:53Z, #2092 21:44→00:04Z) whose rejections CLUSTER at ~10:13–10:33Z and ~22:03–00:04Z — consistent with Thomas batch-clearing pending permission dialogs twice. Two more mechanism-relevant facts: (a) choom=failed (sudo -n DENIED) appears in sibling detached fits the same day — the sudo -n choom prelude is not passwordless for these sessions; (b) the WORKING shape was demonstrated in-fleet at 16:48Z: session a4ecf84a (#1992) wrote the identical gate to /tmp/issue-1992-lint-gate.sh and launched it via a plain bash call — no prompt, gate passed, PR #1770 merged. A third failure shape also fired: a plain background-Bash gate launch died WITH its session (no verdict file for 2h35m — the harness kills session-tied bg Bash), so plain bg-Bash is not the fix either.

unverified hypothesis — verify at plan time: the dominant mechanism is a PERMISSION PROMPT on the inline sudo -n choom … rm -rf compound (supported by the rejection clustering, the no-prompt success of the script-file shape, and the absence of any sudo/choom allowlist rule); a PreToolUse hook hang remains a candidate for the no-ack background-dispatch subset (no hook carries a timeout; probed). Neither mechanism was reproduced under control.

Proposed change

Three prongs, planner to confirm ordering:

1. Gate dispatch shape: pin the Step 10d lint-gate recipe in .claude/skills/issue/SKILL.md to the shape that worked (write the gate to /tmp/issue-<N>-lint-gate-<slug>.sh, launch setsid-detached per the existing

   15-min detached-phase rule, poll the verdict file) — never the inline sudo/rm -rf compound, never session-tied bg Bash; and either drop the sudo -n choom prelude (it is denied anyway) or add a scoped allowlist entry for it in .claude/settings.json.

2. Hook timeouts: add explicit timeout fields (e.g. 30–60 s) to every PreToolUse/PostToolUse hook entry in .claude/settings.json so a hung hook fails the call loudly instead of wedging it forever.
3. Watcher fast lane: teach scripts/autonomous_session_watch.py (or scripts/tick_triage.py) to treat "assistant turn ends in a tool_use with no tool_result for > ~20–30 min" as a wedge trigger — today's ~2.2 h stall fence was the only exit and respawns sometimes re-wedged into the identical pending call. Root-cause alongside #1994/#2039 (the lint gate is also slow/hangy under load: 540 s SIGTERM rc=143 observed twice in-session).

Provenance

• fingerprint: cc0ade222675

• workflow_fix_target: .claude/settings.json, scripts/autonomous_session_watch.py, .claude/skills/issue/SKILL.md

• origin: /daily 2026-08-05 problem sweep — miners 1/2/3/4/5 independently converged on this cluster (12+ pending-call wedge events, one per session transcript, no echoes).