crew-retro

Run by the lead. $CREW is ~/.claude/agent-crew (Windows: $HOME\.claude\agent-crew). Edit only .claude/team/ files other than PROJECT.md yourself; a change to PROJECT.md (or any project file) is a coder job. Be short: this is housekeeping, not a report.

1. Measure

Run the ledger for the period since the last retro (read last_retro in roster.md; if it says never, use the whole history):

powershell -NoProfile -File "$HOME\.claude\agent-crew\scripts\ledger.ps1" -Project . -Since <date> -Format md

Add -Alias old=new for agents that were renamed. The ledger gives, per agent type, model and effort, the spawn counts and the share of usage (each row's cost weighted by model price, divided by the total; this is what matters under subscription limits). Report shares by default. Add -Mode cost and report dollars only if PROJECT.md or USER.md says the user pays per token. A warning "unknown model family" means a model the price file does not know: tell the user to add its family to ~/.claude/team/prices.local.json (README, Models). Remember: subagent_tokens in notifications is a context size, not spend; only the ledger is spend.

Also read the lead's own rows (main session, not subagents) in the ledger output or the session transcript usage: the typical context size per turn and how often it was compacted. Report both in one line. If the evidence says the window is wrong (contexts stay far below it and never compact: lower it; compacted several times in a short span and lost state: raise it; stay within 100000-400000), edit .claude/team/session.json yourself (lead_compact_window, reason, updated) and tell the user in one line. No pull request and no approval: the session-start hook copies the value into .claude/settings.local.json at the next session start. If the user set their own value (env var, /autocompact, user settings) the hook leaves it alone; check-team says which wins.

2. Update the evidence

• routing.md "Evidence": add one line per useful measurement or finding (what was tried, cost, result). Change a routing row only if the evidence supports it, and tell the user in one line.
• roster.md: set each agent's Uses from the ledger spawn counts; set trial agents with 3 or more uses to kept or flag them for retirement.

3. Prune under the caps

Caps: state.md 50 lines, routing.md 150, roster.md 80. For each file over its cap, or holding old items: move knowledge to its cold home instead of deleting it.

• finished work: delete from state.md;
• hazards and conventions that are stable: PROJECT.md (a coder job);
• how the project works: the project's own docs (a coder job);
• agent-specific rules: .claude/team/roles/<agent>.md;
• old evidence lines: merge several into one rule under "Learned rules", drop the rest.

4. Team changes

Look at the roster against the ledger. Propose, do not do:

• hire: via the crew-hire skill (same 300+ word paste in 3+ briefs, or a failure class that keeps slipping through);
• retire: via the crew-retire skill (0 spawns in 5 merged features or redundant). Ask the user for a one-line OK for each.

5. Generic lessons

List lessons that are not about this project (rules about delegation, cost, briefs, hooks, models). Compare with $CREW/docs/RULES.md and skip ones already there. If there are any, propose an upstream pull request to exaltedo2/agent-crew: show the user the lessons stripped of costs, names, paths and any project content (one line each). After a one-line OK, hand it to the coder with that text as the task (a branch in a clone of agent-crew, never main). If the user says no, drop it.

6. Finish

Write the current ISO timestamp into .claude/team/.last-retro and into roster.md last_retro:. Run check-team.ps1 -Project . and fix violations. Tell the user in at most 8 lines: cost headline (top agent types), what you pruned, what you propose.