🔍 Agent Review

Octo’s thin wrapper around the shared agent-review skill in openclaw-hub.

  • Shared implementation: ~/git/openclaw-hub/skills/agent-review
  • Octo-specific report output: ~/git/octo/agents/root/reports/agent-review
  • Issue repo: JeffSteinbok/octo
  • Delivery target: #octo-status (the cron job’s final reply)

What lives where

Shared in openclaw-hub

  • scripts/agent_review.py — trajectory + memory scanner, fingerprint dedup, issue filing
  • SKILL.md — full CLI reference, output schema, synthesis steps, Discord format
  • README.md

Local to Octo

  • Report output under agents/root/reports/agent-review/
  • The agent-review-weekly cron job
  • Issue target (JeffSteinbok/octo) and Discord DM target

Synthesis rules

These rules reinforce the grounding constraint from the shared SKILL.md. The root agent must follow them when composing the weekly delivery message.

  • Surface only findings traceable to tool_errors, cron_errors, memory_flags, or source_health.issues in the JSON output.
  • If those fields are all empty/minimal → send a short “clean week” note. Do not fill the absence of real findings with invented suggestions.
  • Never recommend a schedule change unless the JSON contains a cron_job_schedules block with the job’s actual schedule expression. The script currently never emits that block, so schedule suggestions are always prohibited until a future enhancement adds them.
  • Never state or estimate run counts or dollar costs (e.g. “ran 73x/week”, “saves $0.87/week”) — the script has no such data.
  • cron_stats.total_cron_sessions counts session-level errors, not per-job run frequencies. Do not use it to infer how often a named job runs.
  • source_health entries (e.g. “WARNING: 0 events extracted”) describe scan quality, not operational metrics. Do not derive cost/schedule claims from them.

How Octo runs it

  1. Run the extraction script:

    (Command omitted: it runs on the Octo host.)

  2. Read recent memory files — scan files newer than 7 days from:
    • ~/git/octo/agents/main/memory/
    • ~/git/octo/agents/root/memory/

  3. Synthesize and deliver — follow the synthesis and Discord DM format steps in the shared SKILL.md.

  4. Do not auto-commit from cron — leave report files on disk for manual review.

Decision-model check (log-only)

An hour after this job, agent-review-labels (scripts/agent_review_gate.py) asks the decision model which kind each newly filed [Agent Review] issue is, and whether it duplicates the most similar-titled issues. It only logs, for the Monday decision-model check (docs: Decision Models); labels and issues are left as you filed them.

Source: openclaw-hub