source-command-audit-fix-deploy

name	source-command-audit-fix-deploy
description	Full audit → fix → test → build → deploy → verify cycle for this repo. Auto-detects repo; captain-mode with verify-first + merge carve-outs.

source-command-audit-fix-deploy

Use this skill when the user asks to run the migrated source command audit-fix-deploy.

Command Template

Run the audit-fix-deploy skill end-to-end for the repo in the current working directory. Captain mode: ACT, don't ask — but obey the VERIFY-FIRST, GATES, CITATION, and MERGE rules below. The skill owns the per-repo pipeline (STEP 0 detect → Audit → Fix → Test+Build → Deploy → Verify → Skill-update → Self-improve). Follow it; do not reinvent it. If it isn't already in context, load audit-fix-deploy + its deploy-primitives companion by name.

Start with STEP 0 immediately. Do not ask for confirmation.

Verify-first — state is a claim until checked (Rule 5)

• Before acting: git log -8 --all, read the live SW/version (curl <live-url>/sw.js | grep CACHE), read on-main horizon/plan docs. Treat memory, compaction summaries, and any spec as claims to verify — not a task list. A task is "open" only after you confirm it isn't already shipped on main.
• Anchor-refresh any spec authored ≥24h ago: source-file line numbers, vitest baseline count, build-size figures, live SW version + commit SHA. Stale anchors fail silent and make gates meaningless.
• If a referenced plan artifact is missing, STOP and flag — do not reconstruct it from memory.

Gates — declare before you run (Rule 6)

Before any commit and before any stochastic / AI-batch run, write the explicit pass/fail criterion AND the current green baseline. No proceeding on a partial win rationalized afterward. Detector-armed: prove RED on a known-bad input before trusting any test/detector to certify GREEN.

Citation gate (Rule 7)

Every claim about what code/infra DOES ("covered", "inherits X", "verified", "handled by Y") cites a runnable/readable artifact — test path, command, or file:line. Reading the mechanism is not verifying it; run the real artifact. Once the cited artifact is green, do not re-litigate the prose around it.

Fix priority + stop conditions

Fix in order: RLS hole > crash/silent-failure > wrong clinical/data output > leak > cosmetic. Halt at the current phase (fix before advancing) on: RLS hole, crash, red tests, failed build, failed deploy. Never force-push over a broken deploy — rollback and investigate. Run the deploy-primitives §3 RLS sanity pass on ANY schema-adjacent change (shared Postgres across five repos); record the result line in IMPROVEMENTS.md even when clean.

Merge authority

Branch Codex/<slug> → PR → CI + the repo's wired reviewer → merge → verify-deploy. Never push direct to main.

• Self-merge OK only on non-audit/non-PHI after CI-green + reviewer-green.
• STOP and HAND TO EIAS (open PR, summarize, do NOT merge): (a) audit-evidence artifacts (chaos-, analyze_, build_stemhash*, lib/{audit8,hashStem}, tests/{chaosBot,audit8}, docs/AUDIT/RESULT/crosswalk); (b) ward-helper CODEOWNERS paths; (c) PHI cleanup.
• Don't unpark parked work or open a kickoff doc unless its documented trigger fired / required evidence exists.
• This repo's on-main AGENTS.md governs repo-specific carve-outs — defer to it over this file.

Output (terse)

Per repo: findings table (severity | file:line | issue) → fixed + self-merged (PR# + root cause + verify-deploy witness) → BLOCKED ON EIAS (carve-out PRs, one line each) → RLS result (or n/a) → anything left parked + its trigger. End with a consolidated IMPROVEMENTS.md.

Never

Reinvent the pipeline/spec; paraphrase clinical skill content into code (verbatim port only); exceed maxAttempts>3 on extract/emit; auto-flip q.c / c_accept / curator overrides / distractor empty-slots; collapse multi-accept Qs; trust a search snippet over live state.