---

name: workflow-green-repo description: > Drive a repository to a green baseline — typecheck, lint, tests, and build passing from a clean checkout — when fixing is authorized. Use when "make the repo green", "get CI passing", or "fix all the failing tests". license: MIT

workflow-green-repo — Whole-Repository Greening

Degree of freedom: MIXED. Root-cause of each failure [HIGH freedom]; gate discovery, worklist, and no skip/silence [LOW freedom — run exactly].

Bring the repository to a verified-green baseline: every configured health gate passes from a clean state. This is broad, authorized cleanup — the one place where fixing unrelated pre-existing failures is the goal, not scope creep.

Green is defined by a fresh from-scratch run of every gate, not by the failures you happened to see. Enumerate first, externalize the worklist, fix root causes in batches, and re-run until zero.

How to reason

1. Authorize — whole-repo green was explicitly asked
2. Enumerate — real commands + baseline failures, not assumed
3. Fix-root — no skip, narrow, or ignore
4. Prove — fresh from-scratch run of every gate

Worked example

Authorize: user said "get CI passing on this repo." Enumerate: pnpm typecheck 14 errors in lib/billing; pnpm test 3 failures on date TZ. Fix-root: restore the Invoice discriminant the last refactor dropped; fix TZ in the test clock, do not .skip the suite. Prove: typecheck/lint/test/build re-run from scratch → clean; state file has zero open failures.

Self-critique before reporting

• Authorized — this was not a silent rewrite of unrelated debt
• No fake-green — no .skip, @ts-ignore, or weakened assertion
• Fresh prove — green is a new full run, not the last batch
• Right owner — one plan's leftovers → complete-everything; one pattern → burndown-full; ratchet gaming → audit-gate-logic

When NOT to use this [LOW freedom — do not skip]

• The user asked to finish one plan or feature → use complete-everything.
• The change is one searchable pattern across files → use burndown-full.
• No explicit authorization to touch pre-existing debt → confirm scope first; do not silently rewrite unrelated code.

Phase 0 — Confirm authorization and discover gates [LOW freedom — run exactly]

1. Confirm the user wants the whole repo green, including debt that predates their work. If the surface is narrower, route to complete-everything.
2. Discover the real gate commands — never assume them. Read package.json scripts, Makefile, justfile, CI workflows (.github/workflows/*), AGENTS.md, CLAUDE.md, and README. Identify the exact commands for: typecheck, lint/format, unit/integration tests, e2e (if any), build.
3. Note the required scope for each gate (whole repo vs. per-package in a monorepo) and any environment prerequisites (services, env vars, fixtures).
4. Capture the baseline: run each gate once and record the failing output. This is the worklist source, and it separates real regressions from flakes.

Phase 1 — Enumerate the full failure worklist [LOW freedom — run exactly]

Write .cursor/green-repo-state.md as the durable source of truth:

# Green Repo: <repo/package scope>
Gates: typecheck=`...` lint=`...` test=`...` e2e=`...|none` build=`...`
Baseline captured: <date>

## Failures
- [ ] typecheck: <file:line> — <error summary>
- [ ] lint: <rule> — <file(s)>
- [ ] test: <suite/case> — <assertion / error>
- [ ] build: <step> — <error>

## Flaky / environment-dependent (quarantine, do not fake-fix)
- [ ] <case> — <why it is nondeterministic>

## Discovered while fixing (append, then fix)
- (none yet)

## Intentional exceptions (justify each — needs human sign-off)
- (none yet)

## Evidence
- <gate>: <command> → <result>

State the totals: "N failing gates across M areas. Driving all to green." Mirror the worklist into the native task list for UI visibility; the state file wins after context compaction.

Phase 2 — Fix root causes in batches [HIGH freedom]

Work the list in small batches (one gate area or 5–10 related failures):

1. Read the failing code and its intent before editing. Understand why it fails — do not pattern-match a silence.
2. Fix the root cause. Forbidden: deleting/.skip/.only-ing a test, narrowing an assertion so it cannot fail, @ts-ignore / eslint-disable / any to hide a real type error, broad try/catch that swallows, or blanket snapshot updates that mask regressions.
3. A snapshot or golden file may be updated only when the output change is correct and intended — state why.
4. Tick the item in the state file immediately per batch.
5. Append any newly surfaced failure to Discovered while fixing and close it in this run.
6. If a failure is genuinely flaky or needs a real environment you lack, move it to the quarantine/exception section with a precise reason — never fake a pass. Real environment blockers are the only stop condition.
7. Edit surgically: change what causes the failure, not the surrounding file. Tests you add exist only to pin the fix or to match what the repo already keeps; scratch scripts used to reproduce stay out of the repo.

Do not pause between batches to ask "continue?", and do not wrap up because the context is getting long — compaction will happen on a run this size and the state file carries the worklist across it. Keep going until the worklist is empty; re-read the state file instead of trusting memory.

Phase 3 — Prove green from scratch [LOW freedom — run exactly]

Do not report green until every configured gate passes in a fresh run:

1. Typecheck (whole configured scope) → clean.
2. Lint/format → clean.
3. Unit/integration tests → all pass; quarantined items explicitly listed.
4. E2E (if configured) → pass.
5. Build → succeeds.
6. State file: zero unchecked failures outside justified exceptions.
7. If the repo has CI, note whether the same gates run there and whether any are environment-specific (so local green predicts CI green).

Any new failure or regression → add to the worklist and loop back to Phase 2. For approved-plan or wide-change contexts, invoke completion-judge before the final claim.

Phase 4 — Report [LOW freedom — do not skip]

## Green Repo — report
Scope: <repo/packages>
### Result: GREEN | GREEN WITH QUARANTINE | BLOCKED
### Gates
- typecheck: `<cmd>` → clean
- lint: `<cmd>` → clean
- tests: `<cmd>` → <n> passed
- build: `<cmd>` → ok
### Fixed
- <count> failures across <areas>
### Quarantined / exceptions (each justified)
- <case> — <reason, tracking recommendation>
### Environment blockers (real human/infra gates only)
- none | <blocker + one precise ask>

Leave .cursor/green-repo-state.md in place as the audit trail unless asked to remove it. Suggest committing mechanical fixes separately from behavioral ones if that aids review.

Related

• complete-everything — close one plan's connected scope (not the whole repo)
• burndown-full — one searchable pattern to 100% coverage
• audit-gate-logic — whether the green baseline / ratchet can be gamed
• verification-before-completion — evidence levels this skill proves
• completion-judge — independent verdict before a green claim
• debug-error — root-cause a single stubborn failure