session-retro（会议复古）

Review the full conversation and identify actionable improvements to the project's AI agent configuration and documentation.

Never delegate this skill to a subagent. The retro requires full conversation context.

Steps

Step 1: Audit

• Scan the session for:

- Corrections: where you were corrected or redirected. - Repeated patterns: workflows or knowledge applied multiple times. - Failed approaches: dead ends that future sessions should avoid. - Discoveries: codebase knowledge, architectural insights, or debugging techniques learned during the session. - Missing context: information you had to look up that should have been readily available. - Costly research: topics where significant time or tokens were spent exploring the codebase or external sources. Propose adding results to docs/knowledge-base/ so future sessions start with the answer.

• Every correction must produce at least one proposal. Re-scan the conversation to confirm none were missed.
• Produce proposals only, each with: What, Why (evidence), Where (exact file), Risk.
• Changes should follow the Guidelines

Step 2: Approval gate

• Present every proposal to the user using the multi-choice question tool (one question, multiple: true). Each option label is the proposal ID + short title; each description is a one-sentence summary.
• Every correction received during the session must produce at least one proposal. Do not silently drop corrections.
• No edits before explicit approval.

Step 3: Apply mode

• Apply only approved items, minimally.
• No auto-commit; leave changes staged/unstaged per your normal flow.

Step 4: Verify/report

• Run your repo verification policy and report exact commands + pass/fail.

Guidelines

• Prefer updating existing files over creating new ones.
• Keep skills focused: one workflow per skill.
• Keep docs factual and concise.
• Do not add speculative content. Only propose changes backed by concrete conversation evidence.
• Proposed text must match the target file's style and brevity. In particular, AGENTS.md changes must be minimal (one or two short lines per rule).
• Use .agents/skills/ for all agent conventions, including background knowledge (with user-invocable: false).

Where to propose changes (priority order)

1. Skills (.agents/skills/): new skills for recurring workflows, or refinements to existing skills.
2. Human-readable docs (docs/knowledge-base): codebase knowledge, architecture guides, debugging playbooks, patterns. Anything useful to both humans and AI agents. Terms should be defined in docs/knowledge-base/glossary.md. Other topics should be organized into documents with relevant names. Closely related subjects should be grouped under subfolders with understandable names. When creating or updating a document, use markdown links to refer to the glossary.
3. Always-on rules (AGENTS.md): behavioral refinements to the agent interaction model. Keep changes minimal; this file should stay concise.