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.
- 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.
- 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
--allowedTools Read,Edit,Bash restricts capability. --max-budget-usd caps cost per attempt. --permission-mode acceptEdits auto-accepts file edits within the allowed set.
- Re-run the test to verify. The test's exit code IS the fitness function — no separate sandbox / LLM-as-judge.
- 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) |
1---2name: tdd-repair3description: "Test-Driven Repair: given a failing test, spawn a bounded headless `claude -p` (Read/Edit/Bash only, --max-budget-usd capped) that makes the test pass without modifying it. Use when a failing test should drive the fix to the code under test."4argument-hint: "--repo <path> --test <path> --test-command <cmd> [--max-attempts 1] [--budget 5.00] [--model haiku] [--confirm]"5allowed-tools: Bash6metadata:7 source: https://skillmd.com/skills/ruvnet/tdd-repair8---910Surfaces 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).1112## When to use1314- 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).15- Local TDD workflow — write the failing test first (`tdd-workflow` skill), then run `tdd-repair` to drive the green.16- Regression triage — a previously-green test went red; before opening an issue, spend ~$1 to see if the fix is trivial.1718## When NOT to use1920- **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.21- **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.22- **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.2324## Algorithm2526Implementation: [`scripts/tdd-repair/tdd-repair.mjs`](../../scripts/tdd-repair/tdd-repair.mjs).27281. **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.292. **Spawn `claude -p`** with a focused prompt:30 - Failing test file path (read-only intent)31 - Test command (run only)32 - Hard constraint: do NOT modify the test33 - Hard constraint: do NOT add new dependencies343. **`--allowedTools Read,Edit,Bash`** restricts capability. **`--max-budget-usd`** caps cost per attempt. **`--permission-mode acceptEdits`** auto-accepts file edits within the allowed set.354. **Re-run the test** to verify. The test's exit code IS the fitness function — no separate sandbox / LLM-as-judge.365. **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).3738## Output shape3940```json41{42 "success": true,43 "data": {44 "repaired": true,45 "attemptsTaken": 1,46 "mode": "test-driven",47 "before": { "passed": false, "exitCode": 1 },48 "after": { "passed": true, "exitCode": 0, "durationMs": 4321 },49 "attempts": [50 { "attempt": 1, "claude": { "ok": true, "durationMs": 38421, "usage": { "cost_usd": 0.0234 } }, "verify": { "passed": true } }51 ],52 "totalCostUsd": 0.0234,53 "budgetUsd": 5.0,54 "budgetExhausted": false,55 "shape": { "repo": "...", "test": "...", "testCommand": "...", "maxAttempts": 1, "model": "haiku" }56 }57}58```5960## Exit codes6162| Code | Meaning |63|---|---|64| 0 | Test green after repair (success) |65| 1 | Test still red after `--max-attempts` |66| 2 | Config error (test file missing, test already passes, `--no-test-oracle` unsupported, etc.) |67| 3 | Claude CLI exited non-zero (infrastructure failure) |68| 99 | Reserved for safety tripwire (per ADR-153) |6970## Safety posture7172| Layer | Mechanism |73|---|---|74| Cost cap | `--max-budget-usd` default $5, divided across `--max-attempts`. Hard ceiling — claude exits when reached. |75| 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. |76| 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`). |77| Confirmation gate | `--confirm` REQUIRED — without it, returns dry-run plan (mirrors `harness-evolve` / `harness-mint` convention). |78| Hard timeout | 15 min total wall-clock; per-attempt budget of `timeoutMs / maxAttempts`. |79| Pre-flight | Refuses to run if the test already passes (catches `--test-command` typos). |8081## Inspiration8283Modeled on the **Test-Driven Repair** mode from [agent-harness-generator/packages/darwin-mode](https://github.com/ruvnet/agent-harness-generator/tree/main/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:8485- The test command IS the fitness function — no need for variant scoring86- `claude -p` is already in our stack — no new optional dep87- Bounded cost / capability are first-class flags88- Resumable via `--session-id` if iteration is needed8990Conformant mode (no test, write own repro via MCTS) is deferred to a future ADR.9192## Example9394```bash95# Smoke / dry-run (no --confirm yet)96node plugins/ruflo-testgen/scripts/tdd-repair/tdd-repair.mjs \97 --repo /path/to/myrepo \98 --test tests/auth.test.ts \99 --test-command "npx vitest run tests/auth.test.ts"100101# Actually repair (Haiku tier, $5 budget, 1 attempt)102node plugins/ruflo-testgen/scripts/tdd-repair/tdd-repair.mjs \103 --repo /path/to/myrepo \104 --test tests/auth.test.ts \105 --test-command "npx vitest run tests/auth.test.ts" \106 --confirm107108# Bigger model + more attempts for harder bugs109node plugins/ruflo-testgen/scripts/tdd-repair/tdd-repair.mjs \110 --repo . --test tests/regression-2456.test.ts \111 --test-command "npm test -- tests/regression-2456.test.ts" \112 --model sonnet --max-attempts 3 --budget 15.00 \113 --confirm114```115116## Cost ladder117118| Tier | Model | Per-attempt typical | Use when |119|---|---|---:|---|120| 1 | Haiku | $0.02 – $0.20 | First try — most "make red green" bugs are tactical |121| 2 | Sonnet | $0.30 – $2.00 | Haiku failed, or the bug has multi-file scope |122| 3 | Opus | $1.50 – $8.00 | Sonnet failed — architectural reasoning required (rarely worth it for a single failing test) |