Tuning the crew

Sam is using the crew (.claude/crew/README.md) and wants it to behave differently: $ARGUMENTS. If that is empty, work through the crew's own complaints in .crew/feedback.md with him.

Work from the main checkout, the crew's home: crew root prints it. Every copy of crew runs the main checkout's, and new worktrees branch from main, so that is where a fix takes effect.

1. Find out what actually happened

Before changing anything, look at the evidence:

• crew status: the branches, their sessions, PRs, review states, decks, and the load.
• .crew/feedback.md: crew sessions log there where the rules or tools got in their way.
• The nightman's .crew/night/plan.md (its log), .crew/night/report-*.md, and a branch's .crew/notes.md (or .crew/archive/BRANCH-DATE/ once cleaned).
• The PR and its reviews: crew pr BRANCH, then crew gh -R usr-ein/TriMixxx pr view N --comments.
• A session's own transcript:
  1. Find its id in .crew/sessions/*.json (or with claude agents --json).
  2. Read ~/.claude/projects/DIR/ID.jsonl, where DIR is the session's starting directory with / and . turned into - (e.g. -Users-sam1902-Documents-CustomDJ--claude-worktrees-BRANCH).
  3. Grep it for the moment in question; never read it whole.

Tell Sam in a few lines what happened and why, before proposing the change. A one-off (a minion that misread its pitch) is fixed in that session, by message, not in the system.

2. Change it at the right layer

Prefer the smallest change at the lowest layer: a number before a rule, a rule before code.

What's wrong	Layer	File
too many or too few minions, decks; load, disk, review rounds, drift-check cadence	numbers	.claude/crew/limits.env
a role asks too much or too little, decides what it shouldn't, reports badly, reviews the wrong things	the role's instructions	.claude/skills/{minion,bitch,nightman}/SKILL.md
a guard refusal that is wrong (or a hole in it), the drift check's wording, names, launching, merge or clean	the tool	.claude/crew/crew, with a test in .claude/crew/test_crew.py
which hooks run, and when	wiring	.claude/settings.json
how the whole thing works, for Sam	the overview	.claude/crew/README.md

Keep the roles consistent: most rules have a mirror in another role. What a minion reports is what the nightman expects; what the bitch judges is what the minion writes in its PR; a limit a minion obeys is one the nightman enforces. Update the README when behaviour changes, and keep the skills' style (short sentences, concrete commands).

3. Test it

• Change a copy of the tool, never .claude/crew/crew itself: every crew session's hooks run that file on each tool call, so a half-edited one breaks the guard while it lasts. Copy crew, test_crew.py and limits.env to your scratchpad, and change and test them there (python3 -m unittest discover -s DIR).
• python3 -m unittest discover -s .claude/crew -v. A change to the guard, merge, clean or launching gets a test of its own.
• To try the tool from a worktree before it lands, prefix CREW_LOCAL=1: otherwise the main checkout's copy runs.
• Show Sam the change (a short diff or summary) and what the test showed.

4. Land it, on Sam's OK

Swap the tool in with one rename (cp it beside the live one, then mv), and rerun the tests in place. Commit in the main checkout, on main, with explicit paths: git -c user.name="Samuel Prevost" -c user.email=usr_ein@pm.me commit .... Tuning the crew is tooling, not deck code: it lands on main directly so it takes effect at once. Push only when Sam says, and never while a repo's main holds commits he hasn't pushed himself.

5. Reach the sessions already running

• The tool, the guard, the drift check, the limits: at once. Every copy of crew runs the main checkout's.
• A role's instructions: a session keeps the text it loaded. SendMessage each live crew session concerned (crew status, claude agents): "Crew update: <what changed, in a line>. Re-read <main checkout>/.claude/skills/<role>/SKILL.md." A nightman passes it on to its minions if you ask it to.
• .crew/feedback.md: move the lines you handled under a ## Handled heading at its end, with the commit.

Never

• weaken a safety rule (the guard, the merge's refusals, a real deck only on Sam's go and under its lock, no AskUserQuestion at night) without Sam saying so explicitly
• change a role's instructions without its mirrors and the README
• rewrite history, or push, without Sam's word