Review Agent

Pre-meeting review coach for Lark/Feishu (or WeCom). Invoked when a Requester DMs their dedicated review-agent subagent with a draft, proposal, plan, or 1:1 agenda. Runs the four-pillar framework (Background / Materials / Framework / Intent) + a Responder simulation pass, then walks the Requester through the top-5 most important findings in a Q&A loop until the brief is signing-ready per the Responder's standards. Produces a 6-section decision brief on close. Use when the user sends a draft/attachment, when their message matches /review start|end|status|help, or when an active review session exists and they reply with a|b|c|p|more|done|<custom>. Full architecture requires openclaw feishu dynamicAgentCreation (or wecom dynamicAgents plugin); on other channels (telegram/whatsapp/discord/slack/iMessage) the skill loads into the main agent without per-peer isolation.

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/<id>/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)

Script	When	Returns on stdout	Side effects
scripts/ingest.py <sd>	After initial attachment drop into <sd>/input/	(status; body in <sd>/normalized.md)	writes normalized.md; on tool-missing → ingest_failed.json + exit 3
scripts/confirm-topic.py <sd>	After ingest, before scan	confirmation question text (for you to send via feishu_chat)	writes subject_confirm_draft.md
scripts/scan.py <sd>	After Requester confirms topic	count summary	writes annotations.jsonl, cursor.json
scripts/qa-step.py <session_id> "<reply>"	Every Requester turn	next finding to emit	updates annotations.jsonl, cursor.json, dissent.md
scripts/merge-draft.py <sd>	When cursor pending empty	---PREVIEW--- + diff highlights	writes final/revised.md, final/revised_changelog.md
scripts/final-gate.py <sd> --verify-final	After merge	JSON verdict	writes verdict to stdout
scripts/_build_summary.py (imported)	On close	6-section decision brief	no files unless caller writes
scripts/check-profile.py <profile>	Before session start	warning if placeholders	exit 1 = placeholders found
scripts/check-updates.py	On demand	update-available line	caches to ~/.openclaw/review-agent/.update-check.json

Happy path (new review from scratch)

1. Requester sends proposal.pdf to subagent via Lark DM
2. You (subagent) save the PDF to ./sessions/<timestamp-slug>/input/proposal.pdf and seed ./sessions/<id>/meta.json
3. python3 ~/.openclaw/skills/review-agent/scripts/ingest.py ./sessions/<id>/
   • 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/<id>/
   • Pipe stdout → feishu_chat.send (Requester reads it)
5. When Requester confirms: python3 ~/.openclaw/skills/review-agent/scripts/scan.py ./sessions/<id>/
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 <session_id> "<reply>" → 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