workflow-fix: no-progress tick wedge — heartbeats read HEALTHY, stale-redrive reuses exhausted context

Overview / Motivation

An autonomous /issue session burned 2h16m in a no-progress loop that BOTH watcher lanes are structurally blind to. Observed live on #2054, 2026-08-04T01:12Z-03:28Z; recovered only by manual force-respawn.

Goal

Detect a session that wakes, performs no durable work, posts a heartbeat, and sleeps — repeatedly — and force-RESPAWN it (fresh context) instead of re-driving it into the same exhausted context.

Workflow gap

• Bug observed: #2054's orchestrator emitted four consecutive tick heartbeats (01:52, 02:20, 02:52, 03:20) whose own text reported no state change ("Chain state unchanged since heartbeat #2"; "unchanged. Composing Unit A.1 brief; no state change"). Zero commits, zero status changes, zero non-heartbeat markers across 2h16m. No lane escalated.
• Why it is a workflow gap (two independent blind spots):
  1. scripts/tick_triage.py::compute_issue_verdict decides liveness on marker AGE ONLY — if marker_age_s is not None and marker_age_s <= stale_after_s: return ("HEALTHY", ...). A content-free heartbeat resets the clock exactly like real progress. There is no notion of PROGRESS in the predicate.
  2. When staleness DID trip (heartbeats ~30 min apart vs STALE_S_DEFAULT = 25*60), the verdict is STALE-REDRIVE, which re-drives the SAME session. If the root cause is that session's own context exhaustion, re-driving is futile by construction — it re-attempts the same failing turn. Only a fresh session clears it.
  3. The watcher's existing force-respawn (prompt-wedge) lane keys on isApiErrorMessage: true assistant rows. These turns were NOT api errors — they completed successfully and did nothing. So that lane never fires.
• Root cause of the underlying stall (worth encoding separately): the orchestrator chose to compose a subagent brief with FULLY-INLINED code bodies, explicitly to defeat subagent autocompact thrash — heartbeat #2: "fully-inlined code required to survive micro-scoped subagent thrash". That inverts the standing rule in CLAUDE.md § 429 token-pacing ("Keep subagent prompts lean — pass the PATH to the plan/brief, never inline the body") and exhausted the ORCHESTRATOR's context instead. It self-diagnosed the correct fix at heartbeat #3 ("dispatch with a lean brief pointing at parent references rather than fully-inlined code") and could not execute it — the signature of context exhaustion, where cheap short turns (heartbeats) still succeed but the long composition turn never completes.
• Confidence: high
• verified-at-filing: grep -n "marker_age_s <= stale_after_s" scripts/tick_triage.py -> 1 hit at the HEALTHY branch of compute_issue_verdict (~L974), confirming the age-only predicate; sed -n '950,978p' scripts/tick_triage.py read in full, no progress term present; STALE_S_DEFAULT = 25 * 60 at L113. #2054 heartbeat texts quoted verbatim from task.py view 2054 --json events at 01:52:12Z / 02:20:19Z / 02:52:09Z / 03:20:06Z (2026-08-04).

Proposed change (refine in planning)

1. Add a PROGRESS term to the tick verdict. Track, per issue, a durable progress fingerprint — e.g. (latest non-heartbeat marker ts, HEAD sha on the issue branch, status). If the fingerprint is UNCHANGED across N consecutive ticks (suggest N=3) while markers keep arriving, return a new verdict NO-PROGRESS-RESPAWN rather than HEALTHY or STALE-REDRIVE.
2. Route that verdict to a force-RESPAWN (stop + spawn-issue --auto, preserving the session's auto-approve cap), not a re-drive — the whole point is dropping the poisoned context. Bound it the way the other wedge lanes are bounded: per-issue per-UTC-day cap, episode dedup, kill switch.
3. Consider making heartbeat markers self-declaring (a structured progress: none field) so the predicate does not have to string-match note prose. A heartbeat that admits no progress is the cleanest possible signal and the current skill already writes those words.
4. Separately: add an explicit anti-pattern to the lean-brief rule — "do NOT inline code bodies into a subagent brief to defend against subagent thrash; that converts a subagent context problem into an orchestrator context problem. Pass paths; if a subagent thrashes, micro-scope its work instead."

Scope / surfaces

• Primary: scripts/tick_triage.py (the verdict predicate), scripts/autonomous_session_watch.py (the respawn lane).
• Secondary: .claude/skills/issue-tick/SKILL.md (heartbeat shape / structured no-progress field).
• Secondary: CLAUDE.md § 429 token-pacing (the lean-brief anti-pattern in item 4).

Constraints / invariants

• Do not weaken the existing HEALTHY path for genuinely-working sessions that post infrequent markers — the progress fingerprint must key on durable evidence (commits / status / non-heartbeat markers), not on marker cadence.
• Respawn must be bounded and kill-switched, matching the conventions of the existing wedge lanes.
• scripts/workflow_lint.py passes; mapped tests pass.