AGENT SKILL

retro

Retrospective that improves the agent's environment (checks, standards, navigation), not the code. Use when a build finishes, a session went sideways, or the user asks for a retro.

License 许可
MIT
所属仓库
navaniharsh31/keel
最近更新
2026-10-21
Stars
★ 0
Forks
0

01 前言 Front Matter

name
retro
description
Retrospective that improves the agent's environment (checks, standards, navigation), not the code. Use when a build finishes, a session went sideways, or the user asks for a retro.

02 资源文件 1

03 SKILL.md 正文

A retrospective improves the coding agent's environment (the checks, standards and pointers around the code) so future runs go better. The code itself is out of its reach.

Steps

1. Read the primary sources

The current session's conversation, diff since the spec's Base:, review report, every BLOCKED: and red suite; or, if the user names another session, its logs on this machine.

Done when you can list every mistake, detour, and slow lookup in the session.

2. Find candidates

  • Navigation: a file or fact took long to find, or files hide dependencies: a navigation pointer?
  • Automated checks: a check (lint, types, tests, filesystem linters) could have caught a mistake. Read the repo's own check command first (package.json or build-tool scripts, the CI workflow): a check that exists but is unwired or silently broken is the finding, its fix is the wiring, and the code it would flag is a follow-up. A repo with no guardrail (no pre-commit hook and no CI job running its lint/typecheck/test command) is itself a finding.
  • Coding standards: the reviewer agent missed a mistake: add, remove or clarify a rule.
  • AGENTS.md / CLAUDE.md (repo or global) grown large: steering that belongs in standards or checks, and no-ops (instructions that change nothing).
  • Tool economy: an expensive call, or token-hungry custom tooling (CLIs, MCPs), to streamline.
  • Information access: a crucial fact out of reach (teed dev-server logs, read-only third-party access).

Classify each by where its fix lives:

  • Mechanical → a check: a fixed syntactic pattern, banned API, import shape or file-location rule. Prefer the check to a written rule.
  • Judgement → CODING_STANDARDS.md: cross-file consistency, "matches the surrounding style", anything no guardrail could check.
  • Navigation → a CLAUDE.md pointer: one line naming a file and when to reach for it.
  • Out of reach → a follow-up for a new Bug or Change: product code, tests, dependency pins and edits to existing rules.

Done when every item from step 1 is a classified candidate or judged not worth fixing.

3. List, then pick

List the candidates most severe first, each with its evidence, class and exact fix; or say nothing is worth fixing and skip to Done.

Ask through AskUserQuestion on its own, multiSelect, at most 4 candidates per question, the rest in further questions of the same call. Without AskUserQuestion, list and stop: a candidate is applied only when a reply that arrives after the listing names it, and one that names nothing ("take your recommendations") applies nothing.

Done when the candidates are listed and the user has answered every question, or, without AskUserQuestion, the list is printed.

4. Apply

Read WRITING.md, then apply every chosen candidate. Run a new check once: it passes on the current tree and, where reproducible, catches the original mistake. Commit as chore(retro): <what changed>, on the starting branch when the fix branch is already merged.

Done when every chosen candidate is applied and committed.

Done

  • [ ] Every candidate a pick (or, without AskUserQuestion, a later reply) named is applied and committed; nothing else was applied.

04 同仓库更多 Skills