tdd-repair

Surfaces the Test-Driven Repair loop as a ruflo skill. Use when you have a failing test and want the source-under-test fixed automatically, with the test's pass/fail as the verification gate (no LLM-as-judge).

When to use

• Failing CI test from a recent commit — point this at the test file, get a verified fix (or a clear "couldn't repair within budget" receipt).
• Local TDD workflow — write the failing test first (tdd-workflow skill), then run tdd-repair to drive the green.
• Regression triage — a previously-green test went red; before opening an issue, spend ~$1 to see if the fix is trivial.

When NOT to use

• No failing test exists. Conformant mode (--no-test-oracle) is scoped for a follow-up ADR — needs MCTS over repro generation. For now, write a failing test first.
• Architectural changes. This skill is for tactical "make red green" fixes. Cross-module refactors that incidentally break tests should be done by a human or a swarm.
• Untrusted code. The headless claude -p runs with --allowedTools Read,Edit,Bash — no MCP, no network, no arbitrary file writes — but Bash can still read and write the filesystem and reach the network, so allowing it is not a sandbox. Don't point this at code you wouldn't git checkout . after, and run it on a clean git tree.

Algorithm

Implementation: scripts/tdd-repair/tdd-repair.mjs.

1. Pre-flight verify — run the test command. If it already passes, exit 2 (test-already-passes). Repairing a green test is either a no-op or a --test-command typo.
2. Spawn claude -p with a focused prompt:
   • Failing test file path (read-only intent)
   • Test command (run only)
   • Hard constraint: do NOT modify the test
   • Hard constraint: do NOT add new dependencies
3. --allowedTools Read,Edit,Bash restricts capability. --max-budget-usd caps cost per attempt. --permission-mode acceptEdits auto-accepts file edits within the allowed set.
4. Re-run the test to verify. The test's exit code IS the fitness function — no separate sandbox / LLM-as-judge.
5. If green: emit success: true + per-attempt usage. If red after --max-attempts: emit success: false + receipts. Either way, the workspace is left as claude -p modified it (caller can git diff to review).

Output shape

{
  "success": true,
  "data": {
    "repaired": true,
    "attemptsTaken": 1,
    "mode": "test-driven",
    "before": { "passed": false, "exitCode": 1 },
    "after":  { "passed": true,  "exitCode": 0, "durationMs": 4321 },
    "attempts": [
      { "attempt": 1, "claude": { "ok": true, "durationMs": 38421, "usage": { "cost_usd": 0.0234 } }, "verify": { "passed": true } }
    ],
    "totalCostUsd": 0.0234,
    "budgetUsd": 5.0,
    "budgetExhausted": false,
    "shape": { "repo": "...", "test": "...", "testCommand": "...", "maxAttempts": 1, "model": "haiku" }
  }
}

Exit codes

Code	Meaning
0	Test green after repair (success)
1	Test still red after --max-attempts
2	Config error (test file missing, test already passes, --no-test-oracle unsupported, etc.)
3	Claude CLI exited non-zero (infrastructure failure)
99	Reserved for safety tripwire (per ADR-153)

Safety posture

Layer	Mechanism
Cost cap	--max-budget-usd default $5, divided across --max-attempts. Hard ceiling — claude exits when reached.
Capability cap	--allowedTools Read,Edit,Bash — no MCP tools. Bash can still write files and use the network, so this narrows the tool set, it is not a sandbox.
Scope cap	The prompt forbids adding dependencies. The test file is enforced, not just requested: if an attempt changes it, the original bytes are restored and the attempt is scored failed (reason: test-file-modified).
Confirmation gate	--confirm REQUIRED — without it, returns dry-run plan (mirrors harness-evolve / harness-mint convention).
Hard timeout	15 min total wall-clock; per-attempt budget of timeoutMs / maxAttempts.
Pre-flight	Refuses to run if the test already passes (catches --test-command typos).

Inspiration

Modeled on the Test-Driven Repair mode from agent-harness-generator/packages/darwin-mode ADR-175. Key design difference: instead of wrapping metaharness-darwin evolve (population-based search), we drive a single claude -p invocation. Rationale:

• The test command IS the fitness function — no need for variant scoring
• claude -p is already in our stack — no new optional dep
• Bounded cost / capability are first-class flags
• Resumable via --session-id if iteration is needed

Conformant mode (no test, write own repro via MCTS) is deferred to a future ADR.

Example

# Smoke / dry-run (no --confirm yet)
node plugins/ruflo-testgen/scripts/tdd-repair/tdd-repair.mjs \
  --repo /path/to/myrepo \
  --test tests/auth.test.ts \
  --test-command "npx vitest run tests/auth.test.ts"

# Actually repair (Haiku tier, $5 budget, 1 attempt)
node plugins/ruflo-testgen/scripts/tdd-repair/tdd-repair.mjs \
  --repo /path/to/myrepo \
  --test tests/auth.test.ts \
  --test-command "npx vitest run tests/auth.test.ts" \
  --confirm

# Bigger model + more attempts for harder bugs
node plugins/ruflo-testgen/scripts/tdd-repair/tdd-repair.mjs \
  --repo . --test tests/regression-2456.test.ts \
  --test-command "npm test -- tests/regression-2456.test.ts" \
  --model sonnet --max-attempts 3 --budget 15.00 \
  --confirm

Cost ladder

Tier	Model	Per-attempt typical	Use when
1	Haiku	$0.02 – $0.20	First try — most "make red green" bugs are tactical
2	Sonnet	$0.30 – $2.00	Haiku failed, or the bug has multi-file scope
3	Opus	$1.50 – $8.00	Sonnet failed — architectural reasoning required (rarely worth it for a single failing test)