| name | retro |
| description | Turn a bad or long Pawrrtal work session into durable project learning. Use when the user says $retro, asks to learn from a session, says too much went wrong, or asks to create/update skills or rules from recent failures. |
Retro
Use this when the user wants the project to learn from a messy session. The goal is not a report; the goal is to change the next agent's behavior.
Process
- Extract incidents from evidence. Use the current thread, git diff, PR state, logs, and commands. Do not rely on vibes or blame.
- If
.agent/skills/retro/scripts/collect-codex-retro-context.py is missing, use the current conversation as primary evidence and run /root/.agent/skills/retro/scripts/collect-codex-retro-context.py when available.
- Name each failure mode. Phrase it as an action that should happen next time, such as "prove live Telegram with real delivery before claiming it works."
- Choose the durable artifact.
- Repeated workflow mistake: create or update a skill.
- Repo-wide rule: update
AGENTS.md or a .claude/rules/** file.
- Product/code task: create a bean with the
beans CLI.
- One-off fact that should be remembered across sessions: write a Codex memory note only when the user explicitly asked to remember it.
- Prefer small skills over one giant policy. A skill should trigger for a specific situation and fit in one context read.
- Add verification to every skill. The skill must say what evidence proves the agent followed it.
- Reference project skills from
AGENTS.md. Future agents should discover them without guessing names.
- Validate metadata. Every skill description must include clear "Use when..." triggers and stay under 1024 characters.
Incident Patterns To Capture
| Pattern | Durable Fix |
|---|
| Long-running goal drifts into a tiny PR | Add a scope-audit skill that maps requirements to commits, runtime work, PR diff, and missing items. |
| "Works" claimed from a simulation only | Add live-ops verification requiring real browser/channel/provider proof. |
| CI runners consume the wrong disk or wrong trust model | Add runner-ops instructions covering placement, labels, actor gates, and cleanup. |
| Generic core gets provider/channel/tool details | Add extension-boundary instructions and tests at the interface. |
| Secrets or credentials could leak in logs | Add redaction and "do not print" rules to the relevant operational skill. |
| The user asks "why" after a failure | Answer from current evidence first, then make the durable fix. |
Output
When done, report:
- Skills/rules added or changed.
- Which incident each artifact prevents.
- Validation run.
- Anything intentionally left for a follow-up bean or PR.