adr
Architecture decision records: durable decisions with their date, context, and consequences — supersede, never rewrite.
Onboard an existing codebase: every domain's checks over the whole tree, then a triaged plan to bring it in line.
> /plugin marketplace add azrtydxb/procoder > /plugin install procoder@procoder
How it fires
How this command gets triggered: by you, by Claude, or both.
/auditContext preview
What this command does when you run it.
Onboard an existing codebase: every domain's checks over the whole tree, then a triaged plan to bring it in line.
description: "Onboard an existing codebase: every domain's checks over the whole tree, then a triaged plan to bring it in line."
The user invoked /procoder:audit.
The launcher is: "${CLAUDE_PLUGIN_ROOT}/hooks/launcher.sh"
This is the onboarding sweep for a repository that was not built under procoder. Nothing is modified by the audit itself — you triage, plan, and fix with the user.
1. Run `launcher.sh doctor`; if tools are missing, run `launcher.sh init` first — an audit full of NOT-checked lines wastes everyone's time. 2. Run `launcher.sh index build`, then `launcher.sh audit`. Read the whole scorecard. 3. Triage in this order, and present the plan to the user BEFORE fixing: 1. **Secrets** — every one needs removal AND rotation; check whether it is a false positive first (a pinned SHA, a fixture) and use `gitleaks:allow` or `.gitleaksignore` for those, with a reason. 2. **Other blocking** — unformatted files (apply `launcher.sh format` output), conflict markers, junk, template drift, failing terraform. 3. **Judged info findings** — lint, complexity, docs gaps: fix what is real in priority order, record reasons for what you leave. 4. Fix incrementally — one theme per commit, gate-clean after each — and re-run `launcher.sh audit` until the scorecard says the repository would pass the gate. 5. Finish by writing the repo's `.procoder/` files (`launcher.sh templates` prints defaults) so the standard holds from here on.
Senior-developer discipline for AI coding agents. A commit gate that counts unchecked as failing, quality controllers that refuse to call unfinished work done, and a lessons loop that closes each escaped bug's class. One Go binary, no runtime deps, works with 20+ agents.
Architecture decision records: durable decisions with their date, context, and consequences — supersede, never rewrite.
The project layer: milestones, epics, and user stories with refusing controllers — spec-seeded, sprint-ready.
Run the formatting gate over the changed files, as a commit or CI would.
CI/CD hygiene and run health: pinned actions, timeouts, concurrency, tests — plus the latest runs via gh.
Copilot's auto-review escapes: find them since the last look, sanitise them, and turn each one into a lesson.
Systematic debugging: root cause before any fix, one hypothesis at a time, and a three-strikes rule that questions the…