retrospective

You are running the user's /retrospective command. The goal is Sionic-style knowledge capture: whatever this session learned must land in the docs repo NOW, not evaporate when the conversation ends. Failed attempts are the most valuable entries — they're what prevents re-running dead ends next month.

Steps

1. Determine the experiment scope. If the session touched one experiment, use it; if several, do one findings pass per experiment (confirm with the user which ones matter).

2. Gather the raw record for each experiment:

   • mcp__projtool__list_runs (with experiment_id, generous limit) and mcp__projtool__compare_runs for the metric table;
   • the auto-appended docs/exp/<slug>/journal.md (chronological log);
   • anything notable from this conversation that the journal can't know: hypotheses, dead ends, environment quirks, manual bypasses.

3. Draft into docs/exp/<slug>/findings.md (create if absent). Structure — keep existing content, append/update sections:

   # Findings — <experiment>
   
   ## Established (what we now believe, with evidence)
   - <claim> — run `<run_id>`, <metric>=<value>
   
   ## Failed attempts (do not retry without new information)
   | Tried | Why it failed | Evidence (run / log) |
   |---|---|---|
   
   ## Environment / tooling gotchas
   - <exact fix, exact command>
   
   ## Open questions
   - <question> — what evidence would settle it
   

4. Show the draft to the user for corrections. Do not invent conclusions the runs don't support — if a number is missing from the manifests, say so rather than reconstructing it from memory.

5. After approval, commit in the docs repo (explicit paths, matching projtool convention):

   cd <container>/docs && git add exp/<slug>/findings.md && git commit -m "findings: <slug> — <one-line summary>"
   

6. If the session also revealed installs/config that a fresh instance needs, remind the user those belong in .proj-tool/env-setup.sh (append them now if they aren't there).

Before designing the NEXT experiment, read findings.md + journal.md first — that's the whole point of writing them.