blackgirlbytes/debug-tests

Debug Tests Orchestrator

Run a retry loop that fixes failing Playwright tests with fresh worker context.

Invocation

• /debug-tests -> configure + execute
• /debug-tests -in -> interactive mode
• /debug-tests -auto -> unattended mode
• /debug-tests --dry-run -> configure + discover failures + show worker plan only
• /debug-tests --debug -> execute with prompt/worker-trace logging
• /debug-tests --dry-run --debug -> show worker plan + debug log paths only

Core behavior and guarantees

• Main context is orchestrator only.
• Every fix attempt is done by fresh sub-agent worker(s).
• Workers receive focused failure handoff.
• Independent failures may run in parallel.
• Coupled failures run sequentially.

Do not directly apply failure fixes in the orchestrator context.

Configuration inputs (AskQuestion)

At start, ask:

1. mode:
   • interactive
   • auto
2. session behavior (if .cursor/debug-session.json exists):
   • resume
   • fresh
3. worker model tier:
   • default
   • fast
4. max retries:
   • 2
   • 3
   • 5

Derived runtime values:

• MODE
• MAX_RETRIES
• WORKER_MODEL
• STATE_FILE = .cursor/debug-session.json
• DEBUG_MODE (true if --debug is present)

Phase 0 - Session bootstrap

• If state exists and user chose resume, continue from last completed attempt.
• If user chose fresh, reset state and start at attempt 1.
• If no state exists, start at attempt 1.

Phase 1 - Run tests + parse failures

1. Run:

   npx playwright test --project=chromium 2>&1
   

2. Read test-results.json.
3. Extract failed/timedOut results via suites[].specs[].tests[].results[].
4. For each failing result, collect:
   • suite/spec title and test title
   • spec file path + line (if present)
   • status + duration
   • error.message, error.stack, error.snippet
   • attachment paths if present
5. If no failures:
   • report success
   • delete STATE_FILE if present
   • stop

Phase 2 - Classify failures

Classify each failure as:

• independent: different files, or unrelated tests with no shared setup.
• coupled: same describe/setup path, shared fixtures/state, order-sensitive failures.

Prepare worker handoff per failure:

• failing test identity
• file and line
• error summary
• prior attempt history
• likely relevant app/test file paths

Worker prompt template:

You are a focused test-debug worker.

Failure:
- title: [title]
- spec: [file:line]
- status: [status]
- error: [message]
- stack/snippet: [trimmed]
- prior attempts: [history]

Tasks:
1) Read relevant test and app files.
2) Identify root cause.
3) Apply one concrete fix.
4) Return:
   - rootCause
   - fixSummary
   - filesChanged
   - why this fix should resolve the failure

If debug mode is enabled:
5) Write an execution trace to: [debugExecutionPath]
6) Include:
   - ## Received Context (exact worker prompt)
   - ## Files Read
   - ## Root Cause Analysis
   - ## Fix Applied
   - ## Validation Reasoning
   - ## Final Output

Debug paths per worker (when --debug):

• debugPromptPath: .cursor/debug-logs/[timestamp]-attempt-[N]-[failureSlug]-prompt.md
• debugExecutionPath: .cursor/debug-logs/[timestamp]-attempt-[N]-[failureSlug]-execution.md

If --debug is present, orchestrator must:

1. Ensure .cursor/debug-logs/ exists.
2. Resolve each worker prompt with concrete failure details.
3. Write each resolved prompt to [debugPromptPath].
4. Print each resolved prompt in chat before worker launch.

Phase 3 - Dry-run behavior

If --dry-run is present:

• do not spawn workers
• print mode, retries, worker model, failures, batching strategy, and debug paths
• end with: No workers were launched. Remove --dry-run to execute.

Phase 4 - Fix loop with workers

If not dry-run:

1. If --debug, ensure .cursor/debug-logs/ exists and write/print resolved prompts before each worker launch.
2. Spawn workers for independent failures in parallel batches (max 4 concurrent).
3. Process coupled failures sequentially (one worker at a time).
4. Store fix summaries in session history.

Phase 5 - Re-run + mode control

1. Re-run:

   npx playwright test --project=chromium 2>&1
   

2. Increment attempt counter.
3. Update STATE_FILE with retries, attempts, failures, fix summaries, and remaining failures.
4. Re-read test-results.json and compare with previous attempt.

Stop conditions

1. all pass
2. attempt >= MAX_RETRIES
3. no-progress guard (same failure set repeats)

Interactive mode prompt

In -in mode, after each failed attempt:

• continue -> next interactive attempt
• auto -> switch to unattended mode
• stop -> end and print final summary

Remind user: To switch models: change the model in the Cursor dropdown, then reply continue.

Phase 6 - Final report

Print:

• attempts used / max retries
• what was fixed per test and attempt
• remaining failures and current error summary
• recommended next steps
• stop reason (pass, max retries, no-progress, or user stop)
• execution shape (parallel groups vs sequential)
• model tier used
• debug log paths (when --debug)

Delete STATE_FILE if session is complete.

Important constraints

• Do not call external LLM APIs.
• Do not create custom Node helper scripts for parsing failures.
• Use Playwright JSON output (test-results.json).
• Preserve fresh worker context on every retry.
• If interrupted mid-attempt, resume from lastCompletedAttempt and re-run tests before continuing.