feature-retro

name	feature-retro
description	Run a retrospective on recent work — typically on the develop branch, weekly. Analyzes recent commits, current diff, and conversation patterns to suggest improvements to guidelines, skills, architecture tests, or hooks. Triggers when the user says "retro", "let's review", or invokes /retro.
scope	argos

Feature Retrospective

1. Determine the scope

Default scope on develop: all commits since the last archive entry under .ai/learnings/archive/, or the last 7 days if no archive exists yet.

If the user is not on develop, ask once before proceeding: "You're not on develop. Run retro for the current branch's range, or switch to develop first?"

Otherwise, capture:

git log <range> --format='%h %s'   # what was done
git diff <range>                    # what changed

Plus the conversation context since the scope started (best-effort from the current session).

2. Analyze for friction patterns

Look for, in order of importance:

A. Corrections — places where I was told "no, do X instead". Strongest signal for missing guideline or convention.

B. Re-tries — same logical task attempted multiple times. Indicates missing context or unclear conventions.

C. Tool failures + manual fix — commands that failed; user fixed by hand. Often a missing skill or wrong default.

D. Implicit domain knowledge — user explained something I should have known. Strong candidate for guideline or project context.

E. Process gaps — repetitive prep/formatting/post-steps the user did manually. Candidate for a new skill.

IGNORE:

• One-off clarifications about user intent (creative/design decisions)
• Things already covered by existing skills/guidelines (check first)
• Single anomalies without a pattern

3. For each finding, propose (write in GERMAN)

The inbox is the user's private notebook — write findings in German, even though this skill is written in English. The user reviews them in German.

Structure per finding:

### [Kurzer Titel]

- **Beobachtung:** [1-2 Sätze, mit Commit-Hash oder Datei-Ref wenn möglich]
- **Häufigkeit:** [einmalig | wiederholt in Session | systemisch]
- **Empfehlung:**
  - Ebene: [Guideline | Skill | Architecture test | Hook]
  - Scope: [universal | argos]
- **Begründung:** [warum genau diese Ebene — Bezug zur existierenden Struktur]
- **Entwurf:**

[konkreter Text-Vorschlag oder Code]

4. Write to inbox (German)

Append all findings to .ai/learnings/inbox.md. Create the file with a header if missing. Group findings under a session timestamp:

## Session 2026-05-12 14:30

### [Finding 1]
...

### [Finding 2]
...

NEVER apply changes directly. The inbox is the contract — user reviews and decides via /review-learnings.

5. Close out

Report summary in English: "N findings logged to .ai/learnings/inbox.md (in German). Run /review-learnings when ready."