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
- Define the intended outcome and completion standard.
- Extract concrete evidence from inputs, actions, errors, changes, verification, and final output.
- Identify the smallest meaningful Cases.
- Give each Case one proposed Verdict:
PreservePromote to SkillExtract FactorKeep as HabitFix EnvironmentReject PatternOpen Case
- Separate product defects, Skill defects, environment defects, evidence gaps, and user preference.
- 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:
🧙 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.