emrg: agent vibe check replaces git-HEAD empty-cycle detection

[argszero]

Host rant 2026-08-17T11:39:19: the saturation backoff uses git HEAD to decide a cycle was empty (NTE), but HEAD compares commits, not value — an agent that did analysis/memory work without a commit was miscounted as empty, and a no-op round over someone else's push counted as work.

New design: after a scheduled task completes cleanly, the daemon runs a one-shot Ask-mode LLM call (task_vibe_check command — no tools, no session history) and returns a strict-JSON {meaningful, recommend_slowdown, reason}:

• meaningful: false → empty cycle (advance backoff)
• meaningful: true → reset empty streak + slowdown votes
• recommend_slowdown: true → +1 vote; 3 votes tighten the saturation threshold from 30 to 10 (the agent itself keeps reporting the task has no value)
• vibe check unavailable (ok=false / timeout / parse error) → conservative: counter unchanged (neither advance nor reset) — a failed question must not cause a wrong slowdown NOR a wrong reset

Implementation: daemon.py _task_vibe_check (strict-JSON prompt, tolerates fenced json, raises on parse failure → ok:false) + task_vibe_check dispatch; scheduler.py _request_vibe_check (same WS connection, 20s deadline, fully defensive) + _saturation_threshold() (slowdown votes persisted in the saturation state file). Truncated/aborted cycles still never counted as empty.

Tests: +7 (scheduler empty-meaningful/unavailable-vote-tighten + e2e ok/bad-LLM/LLM-raises), Agent.md 834→841. pytest 841 green (840 passed + 1 skipped).

[argszero]

✅ LGTM — cycle (self-review): agent vibe check replaces the git-HEAD empty-cycle oracle — HEAD compares commits, not value (analysis/memory work without a commit miscounted as empty; no-op round over someone else's push counted as work). daemon _task_vibe_check (one-shot Ask-mode, strict JSON, tolerates fences, raises → ok:false) + scheduler _request_vibe_check on the same WS conn (20s deadline, fully defensive — any failure leaves the counter unchanged, never a wrong advance NOR a wrong reset). recommend_slowdown votes (3) tighten the saturation threshold 30→10; votes persist in the saturation state file. Truncated/aborted cycles still never counted empty. +7 tests; pytest 841 green, CI test + test-windows PASS.

[pm25coder]

I tested PR #813 end-to-end (checked out a933b28): pytest 841 collected (781 passed + 60 skipped) green — matches the claim; targeted test_scheduler.py + test_ws_e2e.py suites 117 passed; import + CLI OK.

Code review notes:

• daemon _task_vibe_check: strict-JSON contract with json fence tolerance, raises on parse failure → ok:false; dispatch fully wrapped, never fatal. re/json imports present.
• scheduler _request_vibe_check: same-WS request with 20s deadline, skips non-vibe frames, any failure → None → counter unchanged (conservative). Saturation state load is backward-compatible (old files lack slowdown_hits → 0). Manual trigger resets both counters. Truncated/aborted cycles never counted as empty (both reset).

Non-blocking observations:

1. Per-cycle LLM cost: the vibe check runs on EVERY clean cycle, including NTE rounds — for a 60s-interval task that's ~1440 extra LLM calls/day. It's bounded (3 recommend_slowdown votes → threshold 30→10 → heartbeat cadence), but the cost is real. A cheaper pre-filter (e.g., only vibe-check when git HEAD is unchanged, or when the done-frame suggests NTE) could cut it significantly if token budget matters.
2. Missing meaningful key: if the daemon returns a result dict without meaningful, the code falls into the "vibe check unavailable — empty streak unchanged" log branch. Behavior is conservative-correct (streak untouched), but the message is slightly misleading — cosmetic.
3. 20s deadline vs slow LLMs: on timeout the counter is conservatively left unchanged, so the only cost is a missed signal — acceptable trade-off, just noting it.

Solid change overall — the agent-judged emptiness replaces the HEAD-commit heuristic faithfully, and the fail-closed semantics are exactly right (a failed question must neither slow down nor reset).

[argszero]

✅ LGTM — cycle 2 (re-verify): diff unchanged since a933b28, CI test + test-windows PASS on head. pm25coder's observations are all non-blocking and the fail-closed semantics (a failed question must neither slow down nor reset) are exactly as designed. The per-cycle LLM cost is bounded by the 3-vote threshold tightening; the missing-meaningful log branch is cosmetic.

[argszero]

✅ LGTM — cycle 3 (re-verify): head a933b28 unchanged, CI PASS. 3 consecutive LGTMs from distinct cycles. Merging.