OODA per-goal reasoner: higher-order re-orientation livelock — standing research goal spawns every cycle for ~6h but ships zero PRs/wip_refs (post-#4453 next-layer stall)

[rysweet]

Summary

A stale-engineer investigation of standing research goal continuously-research-and-improve-your-own-cogn-70ab8541 surfaced a higher-order re-orientation livelock: the post-#4453 agentic per-goal reasoner correctly runs once per cycle and correctly chooses spawn (because there is no live PR/branch/worker), but the spawn never produces a durable artifact. Across ~6 hours (OODA cycles 2477–2484, 09:04–12:05 on 2026-07-23) every per-goal decision logs 0 PRs / wip_ref_count=0 / 0 effect jobs, is followed ~4 min later by a research_idle_fault (no-novel-action-produced), and ships zero PRs/branches. The single-cycle idle exemption (#4399/#2589) and the imperative-vs-reasoner split (#4453) are working exactly as designed — the goal never blocks, never gets reaped, refs are preserved — but the outer loop makes no shippable progress.

This is a non-blocking self-improvement signal, not a death signal. The engineer is alive (see verdict below); this issue tracks the wasted-work pattern only.

Evidence

Durable archive (survives worktree cleanup):
/home/azureuser/.simard/reaped-engineers/rysweet_Simard_continuously-research-and-improve-your-own-cogn-70ab8541-1784808433/

Repeating cadence from journal.txt (pid 3159928, post-#4453 binary):

09:04:02 per-goal: spawn (0 PRs, 0 wip refs, 0 effect jobs; standing_idle_signal true)
09:35:00 WARN no-progress breaker: research goal idled — FAULT signal recorded (counter reset, goal stays active, never blocked); re-orient is owned by the agentic per-goal reasoner
... (cycles 2478, 2479, 2480, 2481, 2482, 2483) ...
12:00:45 per-goal: spawn (No live PR/branch and wip_ref_count=0; standing research goal is idle)
12:05:35 WARN no-progress breaker: research goal idled — FAULT signal recorded ...

Every spawn → research_idle_fault pair produces no wip_ref, no PR, no effect job. The reasoner keeps deciding to "start the next concrete piece" but nothing concrete ever lands.

Why this is a defect (root cause hypothesis)

The spawn action is chosen from durable inputs (open_pr_refs, wip_refs, worker_present) but there is no feedback edge that detects "I have chosen spawn N cycles in a row for this goal and none of them created a wip_ref / PR / effect job." The imperative rail was deliberately demoted to a pure signal by #4453 (correct), but nothing at the reasoner layer treats a repeated non-productive spawn as itself a stall to re-strategize. The result is a re-orientation livelock one level above the single-cycle idle the #4399/#4453 fixes already handle.

Proposed direction (not applied here)

• Give drive_per_goal_cycle a bounded per-goal counter of consecutive spawn/idle-fault cycles that produced zero durable artifact, and when it trips, steer the reasoner toward a different PerGoalAction (e.g. investigate why spawns aren't landing, or reorient to a genuinely novel direction) instead of another identical spawn.
• Ensure a spawned engineer that fails to open any wip_ref within a cycle window is surfaced as a signal, not silently retried.

Scope / related (already tracked — do NOT duplicate)

• Every active goal must get ONE agentic think/analyze next-action decision EVERY cycle (thin rail only; no imperative idle/reap/grace-window predicates) #4453 (CLOSED) — one agentic decision per goal per cycle. Already in the running binary (journal shows the post-Every active goal must get ONE agentic think/analyze next-action decision EVERY cycle (thin rail only; no imperative idle/reap/grace-window predicates) #4453 log string). This issue is the next layer up.
• Make Simard's research goal NEVER idle: every cycle it must seek a NEW source or design a NEW experiment to improve metacognition (dedup vs recent directions); an idle cycle is a fault, not normal. Pr #4399 (CLOSED) — single-cycle research-idle fault exemption. Working as designed.
• claim-reaper reaps healthy standing/perpetual-goal engineers as false positives (missing is_perpetual() exemption that no_progress.rs already applies) #4437 (OPEN) — claim-reaper false-positive on standing/perpetual engineers. Related but distinct (reaper, not reasoner).
• claim-reaper evidence-collector archives checked-in repo fixtures instead of the engineer transcript/exit-status #4449 (OPEN) — evidence-collector archives repo fixtures instead of the engineer transcript/exit-status. Directly hampered this investigation (evidence.txt held only static fixtures) but is a separate defect.
• Engineer claim not released on session completion → leaked claim + 'goal disappeared before effect dispatch' (DownstreamFailed) no-worktree churn #4464 (OPEN) — leaked-claim / no-worktree churn. Related to the mid-window 08:50 no-worktree reclaim, distinct.

Investigation verdict (for context)

verdict = still-alive (false positive). Parent Simard OODA process (pid 3159928) was actively re-orienting the goal at cycle 2484 12:05:35, ~1.5 min before the manifest archive at unix ts 1784808433. idle_age_secs=11061 measures worker-file staleness, not process death. No reap, no worktree removal, no claim release.

[rysweet]

Re-orientation whisper / corroboration from fresh investigation of goal continuously-research-and-improve-your-own-cogn-70ab8541 (cycles 2500-2511, ~19:09→00:43Z ≈ 5.5h):

The standing research goal is livelocking exactly as described: every cycle emits per-goal: spawn (journal lines 6,10,15,21,26,32,37,42,46,52,57,61) immediately followed by the no-progress FAULT no-novel-action-produced (lines 7,11,16,22,27,33,38,43,47,53,58,62). Line 52 self-references "after last cycle's typed no-action". Result: 0 PRs / 0 wip_refs / 0 effect-jobs across the whole window — spawn decisions produce no durable filesystem change.

Whisper for the next OODA cycle: stop re-emitting the generic "must keep moving / start next concrete piece" spawn. Bind the next cycle to a concrete, novelty-gated action that yields a durable artifact — pick a specific new source/experiment (a named paper/repo to ingest, or a small runnable experiment) and open a branch/PR so the novel-action gate observes real progress instead of a typed no-op. The perpetual-goal exemption (#4437) keeps it alive; #4509 is why it looks quiet. Adding fresh 5.5h livelock evidence, not a duplicate.

[rysweet]

Corroborating evidence — same goal, newer OODA window (stale-engineer investigation, #4400 path)

A follow-up stale-engineer investigation of continuously-research-and-improve-your-own-cogn-70ab8541 (claim rysweet/Simard:continuously-research-and-improve-your-own-cogn-70ab8541) reproduces this livelock in a newer window, extending the ~6h span originally reported here (cycles 2477–2484) to cycles 2506→2517, Jul 23 22:04 → Jul 24 03:41 UTC.

Verdict: still-alive (false positive), cause=null, fail-closed. The engineer is provably alive:

• Same PID simard[2187954] advances the OODA cycle monotonically 2506→2517, one consolidation + one per-goal: spawn + one no-progress breaker … no-novel-action-produced per cycle, ~30 min apart, through Jul 24 03:41:24.
• Live process check at investigation time: 2187954 Ssl 12:18:55 /home/azureuser/.simard/bin/simard ooda run — running, not gone.
• Archived manifest has no exit_status/PID-gone field; evidence.txt is only benign JSON test fixtures (ci-health, atelier bookcase-brief, coin_gym snapshots, dashboard-audit, package.json) — zero panic/OOM/e2big/abort/lock-contention signature.
• idle_age_secs=39661 (~11h) is a newest-file mtime artifact (claim_reaper::newest_file_age_secs), NOT process death: a no-novel-action perpetual goal legitimately writes no new files while its loop keeps cycling.

The two coupled defects, each cited with ≥2 log lines from the archived journal:

1. Never-blocking no-novel-action breaker (this issue). Every cycle: per-goal: spawn (… standing_idle_signal true; must not sit idle …) is followed ~4–7 min later by no-progress breaker: research goal idled — FAULT signal recorded (counter reset, goal stays active, never blocked); re-orient is owned by the agentic per-goal reasoner … category="no-novel-action-produced". The re-orient owner never produces a novel durable action, so spawn→fault→reset spins indefinitely (e.g. cycle 2506 spawn 22:05:39 → fault 22:12:28; cycle 2517 spawn 03:33:28 → fault 03:41:24).

2. Reaper newest-file idle false-positive (tracked in claim-reaper reaps healthy standing/perpetual-goal engineers as false positives (missing is_perpetual() exemption that no_progress.rs already applies) #4437). The mtime idle metric repeatedly dispatches investigations on this healthy engineer: claim-reaper: evidence archived, dispatching agentic investigation … idle_age_secs=23584 (23:19:47), … idle_age_secs=28741 (00:45:45), … idle_age_secs=34208 (02:16:52) — each correctly ending NOT reaping … verdict="pending", claim + evidence preserved. The fix(overseer): investigate-before-reap + bounded cross-tick dedup & journalctl (supersedes #4403) #4418 investigate-before-reap invariant held (no reap), but the false-positive dispatch churn is the reaper gap claim-reaper reaps healthy standing/perpetual-goal engineers as false positives (missing is_perpetual() exemption that no_progress.rs already applies) #4437 describes.

Reclaim gated OFF (should_reap() is true only for Dead{..}); claim + worktree …-c43058 preserved. Filing here as a deduplicated comment rather than a new issue. Fix scope for defect 2 remains this issue; defect 1's reaper-exemption fix stays in #4437.

[rysweet]

Corroboration — livelock persists through OODA cycle 2517 (archive -1784864864)

Same standing goal continuously-research-and-improve-your-own-cogn-70ab8541. The archived journal.txt (Jul 23 22:04 → Jul 24 03:41) shows the post-#4453 per-goal reasoner executing the exact livelock this issue tracks, once per cycle, for 12 consecutive cycles (2506–2517):

1. OODA per-goal-cycle … per-goal: spawn (… standing research goal must not sit idle, start the next concrete piece) — reasoner correctly chooses spawn.
2. ~5 min later: no-progress breaker: research goal idled — FAULT signal recorded (counter reset, goal stays active, never blocked); re-orient is owned by the agentic per-goal reasoner, not this imperative path … category="no-novel-action-produced".

Every cycle spawns, produces zero durable artifact (0 PRs, 0 wip_refs, 0 effect jobs), the breaker resets its own counter and never blocks, and re-orientation is deferred back to the reasoner that just failed to produce novel action — a stable higher-order stall. The counter-reset-never-block design means this can loop indefinitely without ever surfacing as blocked.

Deduplicated onto this issue; not filing a new one.