patrol

name	patrol
description	Autonomous observation loop. Time-boxed cycles with ratcheting health score, commits findings to grimoires.
allowed-tools	["Bash","Read","Grep","Glob","Write","Edit"]
user-invocable	true

Patrol — Autonomous Observation Loop

Purpose

The Karpathy Loop for network health. Runs time-boxed observation cycles, ratchets the health score, and commits findings to grimoires. This is the autonomous mode — set it running, walk away, come back to a health trail.

Invocation

/patrol                     # Default: 3 cycles, 5-minute windows
/patrol --cycles 10         # Run 10 observation cycles
/patrol --once              # Single cycle (alias for /observe + commit)

Workflow

Step 1: Load State

Check grimoires/gecko/patrol-state.json for active patrol state:

{
  "status": "RUNNING|HALTED|COMPLETE",
  "current_cycle": 3,
  "total_cycles": 10,
  "baseline_score": 72,
  "best_score": 75,
  "started_at": "2026-03-12T00:00:00Z",
  "last_activity": "2026-03-12T00:15:00Z"
}

If status: RUNNING, resume from current_cycle. If no state file, initialize fresh.

Step 2: Execute Observation Cycle

For each cycle:

1. Run /observe (the atomic health check)
2. Read the emitted observation from observations.jsonl
3. Compare health_score against baseline_score

Step 2.5: Reward-Loop Health (the RLAIHF dimension)

Patrol senses the construct LEARNING loop, not just declared-vs-runtime fit. Read the clew ledgers (~/.loa/constructs/packs/*/LEARNINGS.jsonl; node ~/bonfire/clew-loop.mjs where present) and compare CAPTURE RATE against operator activity:

Condition	Reading
heavy operator activity + pending clews accumulating	loop healthy — the backlog is the drain's (/clew <construct>)
heavy operator activity + near-zero capture	loop UNDER-FIRING — root cause is UPSTREAM of capture: main-loop work isn't routed THROUGH constructs, so corrections never become construct-targeted clews. Surface as drift; the fix is routing (/compose, construct agentTypes), not more capture tooling.
capture without distillation for >2 weeks	teachings rotting in the ledger — surface the ripe constructs

Lesson (clew lrn-20260607-gecko-2f35ed): only ~4 clews existed ecosystem-wide despite heavy operator activity. An under-fed reward loop looks like silence — and silence reads as health unless patrol measures the loop ITSELF as a drift dimension.

Step 3: Ratchet Decision

Condition	Action
health_score > best_score	Update best_score, log improvement
health_score == baseline_score ± 2	Stable — log, continue
health_score < baseline_score - 5	Flag degradation, surface contributing factors
API unreachable	Log as DEGRADED, do not halt

The ratchet only moves up. If the network gets healthier, the new score becomes the baseline. If it degrades, gecko surfaces what changed — but doesn't reset the baseline.

Step 4: Commit Findings

After each cycle:

1. Stage grimoires/gecko/observations.jsonl and grimoires/gecko/patrol-state.json
2. Commit with message: gecko: patrol cycle N — health: XX (delta)
3. Do NOT push — accumulate locally, push on /report or manually

Step 5: Kaironic Termination

Stop when:

• All cycles completed
• Health score stable across 3 consecutive cycles (no new signal)
• API is down for 2 consecutive cycles (no point observing a dead network)
• User interrupts

Update patrol-state to COMPLETE or HALTED with reason.

Step 6: Summary

Output a one-paragraph patrol summary:

gecko | patrol complete | 5 cycles | health: 72→75 (+3) | 2 anomalies surfaced | findings in grimoires/gecko/

Outputs

Path	Description
grimoires/gecko/observations.jsonl	Append-only observation log (one line per cycle)
grimoires/gecko/patrol-state.json	Patrol loop state for resume

Trust Boundary

• Patrol READS: API endpoints, GitHub repos, construct manifests
• Patrol WRITES: only grimoires/gecko/
• Patrol COMMITS: only grimoires/gecko/ files
• Patrol NEVER: modifies construct source, pushes to remote, creates PRs

Constraints

• Time-box each cycle to 5 minutes max (Pi philosophy: constrained windows produce better signal)
• Git-as-memory: the observation log IS the long-term memory, not the conversation context
• If context gets heavy, compact — the JSONL has the truth, not the chat