# Review Agent Session

> Use when the user asks EvoZeus to review an Agent session, task, chat, execution trace, diff, report, or final output and decide what should be preserved, fixed, promoted, rejected, or investigated.

- Skill: `metainflow/review-agent-session` (Agent Skill)
- Install (CLI): `npx get-skill add metainflow/review-agent-session`
- Raw SKILL.md: https://api.skillmd.com/api/skills/metainflow/review-agent-session/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: metainflow (https://skillmd.com/u/metainflow)
- Updated: 2026-11-16
- Page: https://skillmd.com/skills/metainflow/review-agent-session

---


# Review an Agent session

## Required input

Use the session, task, output, files, tool traces, or diff already provided. Ask for missing material only when no evidence is available for the requested judgment.

## Review method

1. Define the intended outcome and completion standard.
2. Extract concrete evidence from inputs, actions, errors, changes, verification, and final output.
3. Identify the smallest meaningful Cases.
4. Give each Case one proposed Verdict:
   - `Preserve`
   - `Promote to Skill`
   - `Extract Factor`
   - `Keep as Habit`
   - `Fix Environment`
   - `Reject Pattern`
   - `Open Case`
5. Separate product defects, Skill defects, environment defects, evidence gaps, and user preference.
6. State the highest-leverage next action and how to verify it.

## Output

Lead with the core judgment. Include only evidence needed to defend it. Do not score the Agent or dump internal schemas.

When a reusable Lesson exists, finish with:

```text
🧙 EvoZeus · 捕捉到一条 Lesson｜<一句话>。要记录下来吗？
```

Do not record it until the user confirms.

## Verification target

- Every conclusion points to observable evidence.
- Each proposed action has a clear owner or route.
- Private data is omitted or generalized.
- The response contains one primary next action.


