audit-and-fix

Run a long-horizon audit-then-implement pipeline: produce a structured audit, cluster findings into PR-sized chunks, and spawn parallel subagents to implement fixes as draft PRs.

Trigger: when the user asks for "audit and fix", wants to clean up a domain, or kicks off an overnight cleanup run. Also triggered directly via /audit-and-fix <domain>.

Arguments: $ARGUMENTS — the audit domain (e.g. "hardcoded business logic", "unused exports in moneyballer/pkg/db", "deprecated Storage API calls"). Required — if empty, ask the user.

Steps

1. Run the audit. Invoke the /create-audit skill. This produces docs/audits/<topic-slug>-<YYYY-MM-DD>.md with Findings and Issues tables. Commit the audit file on a new branch audit/<topic-slug>-<date> and push.

2. Cluster findings into PR groups. Read the audit's Issues table and group findings by:

   • Same file or package → same cluster
   • Same severity AND same fix pattern → same cluster
   • Cross-cutting infra changes (test harness, CI config) → their own cluster

   Produce 4–8 clusters, not 30 single-finding PRs. Each cluster is one PR. Print the clustering plan as a table before dispatching:

   Cluster	Findings	Blast Radius	Recommended Title
   1	3	low	"Remove unused exports in pkg/db"

3. Initialize the status file.

   # Audit-and-Fix Status — <topic> — <YYYY-MM-DD>
   
   Audit: docs/audits/<topic-slug>-<date>.md
   
   | Cluster | PR  | Branch | State   | Last Update |
   | ------- | --- | ------ | ------- | ----------- |
   | 1       | —   | —      | pending | <timestamp> |
   

   Write to docs/audits/<topic-slug>-<date>-status.md.

4. Dispatch parallel cluster implementers. Use the Agent tool with subagent_type: "general-purpose" in a single message with multiple tool calls — one agent per cluster, cap at 6 concurrent. Each subagent receives:

   • The audit file path and the specific findings in its cluster
   • Instructions: create a worktree off main, implement the fixes, write or update tests, open a draft PR with required sections (Summary, Test plan, Spec ID if applicable), wait for CI, address Copilot comments autonomously
   • A strict rule: never merge. Leave the PR as draft when done.
   • Required output: {cluster, pr_url, branch, final_state} where final_state ∈ {ready, ci-failing, review-pending, blocked}

5. Monitor and update the status file every 5 minutes while subagents run. For each cluster, record: PR URL, branch, current state, last-update timestamp. If a subagent fails (worktree disappears, CI times out), mark its cluster blocked and record the reason rather than retrying silently.

6. Post the final summary table when all clusters reach ready or blocked:

   Cluster	PR	State	Notes
   1	#201	ready	0 Copilot comments
   2	#202	blocked	rebase conflict on pkg/foo.go

   Do not merge. The user decides which PRs to merge (use /merge-pr per PR).

Rules

• Never merge. This command produces draft PRs only.
• Never force-push. Never use destructive git operations in subagents.
• If a worktree path disappears mid-run, recreate it — do not fail the whole run.
• Cap parallelism at 6 concurrent subagents. Batch the rest.
• Every draft PR must include the audit file path in its Summary section so reviewers can trace the finding.
• Keep the status file updated even on failures — silence is worse than partial information.
• If clustering produces more than 8 PRs, STOP and ask the user to narrow the audit scope.

Compute scope

The Frontier floor describes the coordinator's consumer requirement. Native skill frontmatter does not enforce that floor. Keep the existing general-purpose worker dispatch, inheritance, tools, worktrees and draft-only merge prohibition. This declaration adds no enforced worker-floor guarantee and does not set a worker's model or effort. Report coordinator and worker observations separately when they are available; an unavailable observation remains unknown.