test-fix-loop-n

---

name: test-fix-loop-N description: Discover Go test failures, partition by package/count, spawn N parallel test-fix-loop agents, verify results. Args: [max-agents]

Test Fix Loop N

Discover all Go test failures, partition across N parallel agents, each running /test-fix-loop. Agents merge back to shared parent branch via existing sync protocol.

Arguments

• $1 — parent branch (required). e.g. feat-xslt3.
• $2 — max agents (optional, default 4).

If no parent branch given, ask user.

1. Setup

1. Confirm $PARENT exists: git rev-parse --verify $PARENT.
2. Create orchestrator worktree:
   • Name: fix-orch-<short-id>-<date> — globally unique.
   • Path: $PROJECT/.worktrees/<branch>.

   git worktree add $PROJECT/.worktrees/<branch> -b <branch> $PARENT
   

3. cd to orchestrator worktree. Create .tmp/ if missing.
4. This worktree is for discovery + final verification only. Orchestrator does NOT fix code.

2. Discovery

1. Run full suite:

   go test ./... -timeout=15m -v > .tmp/go-test-discovery.txt 2>&1
   

   -v required — ensures every test name appears in output.
2. Parse .tmp/go-test-discovery.txt. Extract every failure:
   • --- FAIL: TestName → capture test name
   • FAIL\tpkg/path → capture package
   • panic: / compile error → capture package, mark __compile_error__
3. Build failure manifest: list of (package, test_name) tuples.
4. Store in .tmp/failure-manifest.txt.
5. Zero failures → report clean, stop.

3. Partitioning

1. Group by package → (pkg_path, [test_names], count).
2. Sort by count descending.
3. Compile-error packages → isolated partition each (one package per partition).
4. Target partition size: max 8 tests per agent. Agents have finite context — fewer tests means more context budget per test, which means more actual fixes. Use min(8, ceil(total_failures / max_agents)) as the target.
5. Large packages (count > target) → split into sub-partitions by test name.
6. Small packages → combine into one partition if combined count <= target.
7. Dispatch at most max_agents partitions per wave.
8. If there are more partitions than max_agents, do NOT merge them past the 8-test cap — dispatch the largest max_agents partitions now and defer the rest to the next wave (retry wave re-discovers remaining failures automatically).

Per partition compute:

Field	Value
Package list	One or more ./pkg/path entries
-run regex	^(TestA|TestB|TestC)$ — anchored, metacharacters escaped
Omit -run	When partition owns ALL failures in a package

Store partition plan in .tmp/partition-plan.txt.

Regex Rules

• ALWAYS anchor with ^...$ — Go -run does substring match.
• Escape regex metacharacters in test names (+, ., (, etc.) with \.
• Parent test name suffices for subtests — Go runs all subtests when parent matches.

4. Spawn Agents

Issue ALL Agent tool calls in single block (parallel execution).

IMPORTANT: Subagents do NOT have access to the Skill tool — they cannot run /test-fix-loop. Instead, tell them to Read the skill file and follow it directly. They DO have Read, Edit, Write, Bash, Grep, Glob, and other core tools.

Subagents also run in agent isolation worktrees branched from main — NOT from $PARENT. The prompt MUST instruct the agent to create its own worktree from $PARENT and work there exclusively.

Each agent receives prompt (template — substitute actual values):

You are a test-fix agent. Read and follow the procedure in this file EXACTLY:

  ~/.claude/skills/test-fix-loop/SKILL.md

Arguments for the procedure:
  - Parent branch ($PARENT): <parent-branch>
  - Extra test args: -run '^(TestA|TestB|TestC)$' ./pkg1 ./pkg2
  - Project root ($PROJECT): <absolute path to repo root>

CRITICAL SETUP: Your current working directory is NOT on the correct branch.
You MUST create a new worktree from $PARENT before doing anything else:
  cd $PROJECT
  git worktree add $PROJECT/.worktrees/<unique-branch> -b <unique-branch> $PARENT
Then cd to that worktree. ALL work happens there.

You are responsible for fixing ALL of these assigned failing tests:
1. ./pkg1 — TestA
2. ./pkg1 — TestB
3. ./pkg2 — TestC

CONTEXT CONSERVATION — READ THIS CAREFULLY:
Your context window is finite. If you waste it on analysis, you will run out
before making fixes. Follow this workflow strictly:

1. Run your assigned tests ONCE to find failures.
2. Pick the FIRST failing test. Debug it, fix it, verify it.
3. Run `golangci-lint run ./...` — fix any lint errors.
4. COMMIT and MERGE immediately — do not batch fixes.
5. Sync, re-run tests, pick the next failure. Repeat.

DO NOT analyze all tests upfront. DO NOT read entire files — use Grep to find
specific functions. DO NOT spend context exploring blockers — document and move on.
Every iteration MUST produce a commit or a documented blocker. Pure analysis
with no commit is wasted context.

COMPLETION RULES:
- MUST attempt every assigned test. Do NOT stop after fixing one or a few.
- After each fix+merge cycle, re-check which assigned tests still fail.
  Continue until all fixed or documented as non-actionable blockers.
- Test now passes (another agent fixed root cause) → counts as resolved, move on.
- "Complex" / "requires large change" is NOT a reason to stop. That IS the work.
- DONE only when: (a) all assigned tests pass, OR (b) every remaining failure
  documented as non-actionable blocker with specific reason.

REPORT FORMAT when done:
For each assigned test: test name, status (fixed/already-passing/blocked),
root cause, fix description or blocker reason.

• Each agent creates own worktree (prompt instructs this explicitly).
• Cross-agent fix propagation: existing sync protocol resets to parent HEAD each iteration.
• Wait for all agents to return.

5. Collect & Validate Reports

1. Parse each agent's return: fixed tests, blockers, errors.
2. Validate completeness: every assigned test MUST appear in agent's report.
   • Missing tests → record as agent abandoned — needs retry.
3. Consolidate into .tmp/agent-reports.txt.

6. Final Verification

1. Sync orchestrator worktree to $PARENT HEAD:

   git fetch . $PARENT
   git reset --hard $PARENT
   

2. Run full suite:

   go test ./... -timeout=15m > .tmp/go-test-final.txt 2>&1
   

3. Cross-reference against original manifest → categorize each test:
   • fixed — was failing, now passes
   • blocked — agent documented as non-actionable
   • abandoned — agent quit early, test still failing

7. Retry Wave

This step is AUTOMATIC. Do NOT stop to ask the user. Proceed immediately after final verification if the trigger condition is met.

Trigger: failures remain AND (wave made progress OR any agent abandoned tests).

1. Re-run discovery (step 2) against current state.
2. Collect only tests still failing.
3. Exclude tests confirmed as non-actionable blockers (only truly non-actionable: infra, permissions, network — NOT "complex" or "requires large change").
4. If remaining count > 0 → re-partition and spawn new wave (steps 3–6).
5. No wave cap. Keep spawning waves as long as progress is being made.
6. STOP only if:
   • Wave made zero progress (same subtest failure count as previous wave), OR
   • All remaining failures are truly non-actionable. Progress is measured by SUBTEST failure count, not just top-level test count.

Between waves: check for uncommitted work in agent worktrees. If an agent left uncommitted changes, verify them (run the agent's assigned tests with and without the changes). If they pass more tests → commit and merge before the next wave. This recovers work from agents that ran out of context before committing.

8. Cleanup

1. Remove orchestrator worktree:

   cd $PROJECT
   git worktree remove .worktrees/<orch-branch> --force
   

2. Agent worktrees left per existing convention.

9. Final Report

## Orchestrator Report

Failures discovered: N
Waves: W
Agents spawned: M (wave 1) [+ K (wave 2) ...]
Tests fixed: X
Tests blocked: Y
Tests abandoned: Z

### Per-Agent Summary
| Agent | Wave | Worktree | Assigned | Fixed | Blocked | Abandoned |
|-------|------|----------|----------|-------|---------|-----------|

### Remaining Failures (if any)
| Package | Test | Status | Reason |
|---------|------|--------|--------|

Hard Rules

• ALWAYS discover in orchestrator worktree, not repo root.
• ALWAYS spawn via Agent tool in single parallel block.
• NEVER exceed max_agents.
• NEVER mix compile-error packages with test-failure packages in same partition.
• ALWAYS anchor -run regex with ^...$.
• ALWAYS run final verification after all agents complete.
• Keep running waves until zero progress or all non-actionable.
• NEVER skip retry wave if agents abandoned tests or wave made progress.
• ALWAYS include full test assignment list + completion rules in agent prompts.
• ALWAYS validate agent reports for completeness — detect early quitters.
• ALWAYS re-queue abandoned tests in retry waves.
• NEVER accept "complex" / "large change" as blocker reason — only non-actionable items (infra, permissions, network).

Actionable vs Non-Actionable

Same criteria as test-fix-loop:

Actionable	Non-Actionable
Implementation bug	External infra outage
Wrong error code / serialization	Corrupted checkout
Missing feature needed by test	Toolchain install failure
Flaky test from repo code	Permission / network issue
Parser / fixture issue fixable in repo