| name | go |
| description | Final completion/sanity check — audit the current code with a team of subagents to prove it is really finished (no stubs/TODOs/no-ops), plus correctness bugs and leanness issues, then fix them and verify. Use when the user types /go, or asks to sanity-check completion, audit-and-fix, review-and-fix, clean up, or harden the current branch/diff. |
/go — team completion check, audit & fix
Run this as a sanity completion check: prove the work is really done — not stubbed, half-wired, or TODO'd — and free of correctness bugs and needless cruft. Orchestrate a multi-agent workflow that audits the code under review for (1) completeness (no stubs/placeholders/no-ops masquerading as done), (2) correctness defects, and (3) leanness/cleanliness issues, applies the fixes, and verifies against the project's own gate (lint + type-check + tests). Then independently verify and report.
How to run it
-
Launch the bundled workflow (it runs in the background; you're notified on completion):
Workflow({
scriptPath: "<REPO_ROOT>/.claude/skills/go/audit-workflow.js",
args: { focus: "<anything the user typed after /go, else empty>" }
})
Resolve <REPO_ROOT> with the current working directory (the repo you're in). Pass any words the user added after /go as args.focus to narrow scope (e.g. /go the providers layer).
-
When it completes, independently verify — don't just trust the agents. Re-run the project's full gate yourself (discover it: ruff check . && ruff format --check ., ty check or mypy, pytest with coverage if configured — prefer the project's .venv binaries) and confirm green.
-
Report: the confirmed correctness bugs fixed, the leanness fixes applied, anything deliberately left (with rationale), and the verification results.
-
Do NOT commit or push unless the user explicitly asks. Leave the working tree with the fixes applied and summarize.
What the workflow does (phases)
- Scope — a scout agent uses git to list the files under review (commits on the current branch vs the default branch, plus uncommitted working changes); falls back to the package source if the tree is clean.
- Audit (parallel) — correctness reviewers (logic bugs, edge cases & error handling, types & contracts) and leanness reviewers (dead code & duplication, over-abstraction & complexity, stale comments & cruft). Each returns structured findings with severity, confidence, file, and a proposed fix.
- Verify findings (adversarial) — every correctness finding is independently re-checked by a skeptic; false positives are dropped before anything is touched.
- Fix (file-owned, parallel) — confirmed correctness fixes + leanness fixes are grouped by file so no two agents edit the same file; each agent applies minimal, behavior-preserving changes to its files only.
- Gate — one integrator discovers and runs the project's lint/type/test, fixes any breakage, and reports status + coverage.
Principles
- Safe by default: apply high-confidence correctness fixes and behavior-preserving leanness cleanups. Anything risky or behavior-changing is reported, not force-applied.
- No silent scope changes: if a fix would change behavior or public API, surface it instead.
- Idempotent: if there's nothing to fix, the workflow says so rather than inventing work.
- This skill only orchestrates and fixes; committing/pushing is the user's call.