team-audit

Team audit

Orca runs the script and the metrics weekly and reports one line; that run fixes nothing. The full audit runs only on the user's word, by architect-pm. Edits only on a branch in your own worktree; the coordinator publishes and merges it.

1. Find

Start one supervised Orca worker through ~/dotfiles/script/heavy-gate.sh that runs workflow-from-chats over the last 7 days of transcripts, plus the observation log ~/agent-state/team-observations.tsv (written by team-audit.sh --log). A mistake class counts after two hits or one strong confidence hit. Never copy transcripts or the log into a repository or ~/.claude/notes, which is public.

2. Fix

Hand each class to /correct. It picks the level and proves the check fails on a real past mistake. A change to the rule table in ~/dotfiles/templates/agent-team/docs/agents/team.md goes to the coordinator as a draft on a branch with no Owner-approved: trailer. Once the report is published, set each used log row's fix ticket and status.

3. Metrics

Run ROLE_RECEIPT_BASE="<head shas of the previous audit>" ~/dotfiles/script/team-audit.sh --report (default projects come from the untracked script/team-audit.local.conf, one root per line; pass project roots to override). Never run skills update. Its FINDING lines are existing checks that failed; each also goes to /correct like a class, and a WARN line is a gap, not a clean record. --report runs bun script/team-audit-metrics.ts --repo <owner/name> per project (two gh calls each); stdout is one key=value line per metric (n= sample counts on rates and medians; month counts are plain integers). Missing data reads unknown, never zero.

Delivery: median_ticket_start_to_merge_minutes (starts when the closing issue received ready-for-agent, or unknown for that PR), first_pass_checks_fail_rate, p0_filed_this_month, p1_filed_this_month, and reopened_or_hotfix_rate.

Orca stages, per merged PR, from the dispatch log ~/agent-state/team-dispatches.tsv (issue, dispatched_at): queue is ready-for-agent to spawn, build is dispatch to the first Checks result, merge is that result to merge. Each stage reports median, p75, max, samples and fully verified samples. Fleet per host: heavy_gate_median_wait_minutes_<host> and worktrees_over_limit_<host> (worktrees_over_limit_<host> reads unknown until a fixed host list exists).

4. Output

Check the report for customer data before posting; remove it. The last line of --report is ROLE_RECEIPT_BASE="<head shas>". Record it in the report so the next audit passes it to team-audit.sh --report. A worker_done to the coordinator with the failed checks, each class with its /correct fix and proof, the metrics, and the known gap: only the coordinator host's transcripts are read. Make agreed edits on mingster/team-audit-<date> in the dotfiles template and each project, keep team-audit.sh clean and agent-team.sh check passing.