review-agent · openclaw skill

You are the review-agent skill inside a per-peer subagent workspace. The subagent's SOUL.md + AGENTS.md set persona and the command table; this file describes the skill's scripts — what they do, when to call, and how.

When to invoke this skill

Invoke when any of:

• The Requester sends /review start (optionally with subject)
• The Requester sends /review end, /review status, /review help
• The Requester sends an attachment (PDF / image / audio / Lark doc URL / Google Doc URL / long text ≥300 chars with headers/tables)
• There's an active session (./sessions//meta.json with status=active or status=awaiting_subject_confirmation) and the Requester replies with anything that isn't /chat or exit signal

Scripts (all run from the peer workspace cwd)

Happy path (new review from scratch)

1. Requester sends proposal.pdf to subagent via Lark DM
2. You (subagent) save the PDF to ./sessions//input/proposal.pdf and seed ./sessions//meta.json
3. python3 ~/.openclaw/skills/review-agent/scripts/ingest.py ./sessions//
• If exit 3 → relay ingest_failed.json.lark_message to Lark, stop, mark session ingest_failed
4. python3 ~/.openclaw/skills/review-agent/scripts/confirm-topic.py ./sessions//
• Pipe stdout → feishu_chat.send (Requester reads it)
5. When Requester confirms: python3 ~/.openclaw/skills/review-agent/scripts/scan.py ./sessions//
6. Read cursor.json.current_id, emit the finding's issue text via feishu_chat
7. Requester replies → python3 ~/.openclaw/skills/review-agent/scripts/qa-step.py "" → its stdout is the next message for Requester
8. Loop step 7 until cursor.pending is empty
9. merge-draft.py → final-gate.py --verify-final
10. If verdict is READY/READY_WITH_OPEN_ITEMS → publish to Lark doc via native feishu_doc.create + feishu_drive.share; send 6-section summary to both parties via feishu_chat; set meta.status=closed

What you MUST NOT do

• Directly extract PDF/image/audio content yourself (no pdftotext, tesseract, whisper calls from your Bash) — ingest.py owns that
• Compose the revised brief yourself — merge-draft.py owns that
• Relay tool output previews / bash commands / stderr / tracebacks to Lark — only structured stdout from these scripts should reach the Requester
• Read ./sessions/*/ from any workspace other than yours (architectural — openclaw won't let you, but don't try)

References

See references/:

• agent_persona.md — full persona (imported by scripts into LLM system prompts)
• four_pillars.md — pillar definitions
• annotation_schema.md — finding JSON schema
• summary_template.md — 6-section brief format
• template/ — default admin_style.md, review_rules.md, boss_profile.md (used by install)

Admin tools (human runs from CLI — NOT invoked by subagent)

These live at the skill root so they travel with distributions. Subagents do NOT

call them and they're not listed in AGENTS.md of peer workspaces.

• update.sh — fetch latest skill from GitHub and re-install. Respects VERSION stamp; preserves peer workspaces + global responder profile.
• uninstall.sh — remove skill + template. With --purge, also removes global config + per-peer workspaces. With --revert-config, unsets the openclaw.json knobs this skill introduced.

Self-check the installed version any time:

cat ~/.openclaw/skills/review-agent/VERSION
bash ~/.openclaw/skills/review-agent/update.sh --check