Skip to main content

retro

Run a retro, a review of finished work that turns its friction into ranked, evidence-backed fixes. Use when Joel says retro, $retro, retrospective, "what went wrong", "review the finished lanes", "full project retro", or names a scope to review ("$retro <path or project>"), or after an incident or friction that repeats across lanes. The owner launches a read-only judge lane; the judge follows the procedure in the muster skill's references/retro.md.

Source facts

Repository
joelhooks/pi-muster
Last source activity
October 8, 2026 at 00:49
Detected SKILL.md language
English
Stars
3
Forks
0

Install options

The review-first prompt is selected by default. You can switch to a direct command or download a local copy.

Review the source files

Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.

Showing SKILL.md

SKILL.md
Source instructions · Read-only preview
name
retro
description
Run a retro, a review of finished work that turns its friction into ranked, evidence-backed fixes. Use when Joel says retro, $retro, retrospective, "what went wrong", "review the finished lanes", "full project retro", or names a scope to review ("$retro <path or project>"), or after an incident or friction that repeats across lanes. The owner launches a read-only judge lane; the judge follows the procedure in the muster skill's references/retro.md.
# Retro A retro reviews finished work and returns ranked environment fixes: parked lanes and desk decisions. It never fixes anything itself. Its purpose is to improve the process and the code toward the project's `VISION.md`, so rank candidates by how far they move the project toward it. Retros run only after an incident (a bad landing, outage or lost work), friction that repeats across lanes, or when Joel asks. There is no standing judge slot. ## Triggers The desk opens a retro by hand and records the incident, repeated friction, or Joel's request in the brief. A lane count, elapsed day, or ordinary project review does not trigger one. Keep the scope narrow and off unrelated delivery's critical path. Don't stack retros while one is already running. ## Model The judge runs on `openai-codex/gpt-6-astra` at `xhigh`. When token burn is heavy (a long batch, many sessions, or the project's spend is already high), use `openai-codex/gpt-6.1-sol` at `high` instead, and say so in the brief. ## If you are the owner (a desk or boss) 1. **Scope.** Select the lanes relevant to the incident or repeated friction; `project_review` lists unreviewed closed work as `retroLanes`. When Joel names a scope (a path, a project, or "the whole project"), that scope is the batch. Put the trigger and his words in the brief. 2. **Open a retro.** `lane_open` with `kind: "retro"` and a fresh slug such as `retro-<YYYY-MM-DD>`, even at full WIP. Only one retro runs at a time. 3. **Brief.** Name the scope, the evidence (sessions, packet reports, closed tails, gate logs, retros already done), the questions that matter most, and one absolute artifact path under `.brain/data/muster/retros/`. Tell the judge to follow `../muster/references/retro.md`. 4. **Launch a judge.** `agent_launch` with `role: "judge"`, `cwd` the project dir, and the model from [Model](#model). When Joel names a different model, use exactly that one. 5. **Land and route.** `packet_verify`, then `packet_land committed` (artifact packet), then route each candidate: parked lanes with `lane_open open: false`, decisions with `desk_post`. Then `project_review retro: true`. ## If you are the judge Follow [references/retro.md](../muster/references/retro.md) exactly: bounded session slices, cited evidence, ranked candidates, one artifact packet. Read-only: no edits to code, skills, or steering files.
View on GitHub