Test Lead

Orchestrate browser test execution from a design doc's ## Test Plan section. Spawn test-executor agents one at a time, collect results, update the living state doc, and present a summary.

1. Parse Config

Read .project-orchestrator/project.yml (use defaults if missing):

Key	Default	Validation
test.model	sonnet	Must be haiku, sonnet, or opus — error otherwise
test.screenshot_on_failure	true	Must be boolean — error otherwise
test.base_url	null	If set, must be a valid URL string — warn if not reachable

If validation fails, report the error and stop.

2. Find & Parse Design Doc

Resolution order:

1. If plan file path provided in $ARGUMENTS, use it
2. Otherwise, check {config.plans_dir}/INDEX.md for active plans
3. Otherwise, glob for {config.plans_dir}/*-design.md

No design doc found: "No design doc found. Nothing to test."

No ## Test Plan section: "No test plan found in design doc. Add a Test Plan section first."

Extract from the test plan:

• Base URL from ### App Context
• Setup steps from ### Prerequisites
• Scenario table from ### Test Scenarios
• Scenario details from ### Scenario Details
• Dependencies between scenarios (from Precondition column)

3. Pre-flight Validation

All checks must pass before creating the team. If any fail, abort — no cleanup needed.

1. MCP check — attempt list_pages via Chrome DevTools MCP.

   • If fails: "Chrome DevTools MCP is not available. Please ensure Chrome is running with DevTools MCP configured, then try again." Abort.

2. Base URL — use test.base_url from config as default, or base URL from test plan.

   • Ask user to confirm the URL is accessible.
   • If user says it's not accessible, abort.

3. Setup steps — present any setup steps from ### Prerequisites and ask user to confirm they're done.

4. Write State File

Write .project-orchestrator/state.json (atomic: write .tmp then mv):

{
  "active_plan": "{relative path to design doc}",
  "slug": "{feature slug}",
  "team": "test-{slug}",
  "phase": "testing",
  "started": "{ISO 8601 timestamp}"
}

Ensure .project-orchestrator/ directory exists before writing. If write fails, report error and exit.

5. Create Team

TeamCreate("test-{slug}")

If TeamCreate fails: delete .project-orchestrator/state.json, report error, exit.

6. Create Screenshot Directory

Ensure .project-orchestrator/screenshots/{slug}/ exists:

mkdir -p .project-orchestrator/screenshots/{slug}/

7. Execute Scenarios Sequentially

1. Create tasks — TaskCreate for every scenario in table order (T1, T2, T3, ...). Include scenario title, steps, expected outcomes, and preconditions in each task description.

2. For each scenario, one at a time:

   a. Dependency check — if the scenario depends on a prior scenario that failed, mark it as skipped (reason: "dependency T{N} failed"). Do NOT spawn an agent. Update the living state doc and continue to the next scenario.

   b. Spawn test-executor agent:

   Task(
     subagent_type: "test-executor",
     model: config.test.model,
     team_name: "test-{slug}",
     name: "test-{scenario-slug}"
   )
   

   Prompt must include:

   • Scenario number and title
   • Full scenario details (steps + expected outcomes)
   • Living state doc path
   • Base URL
   • config.test.screenshot_on_failure value
   • Screenshot directory path: .project-orchestrator/screenshots/{slug}/

   c. Wait for agent to complete or go idle.

   d. Collect result from TaskUpdate metadata:

   {
     result: "pass" | "fail",
     steps_total: N,
     steps_passed: N,
     failure_step: "Step 2 — ..." | null,
     failure_reason: "..." | null,
     screenshot_path: "..." | null
   }
   

   e. Update living state doc — mark scenario status and result in the test plan table.

   f. Shutdown the agent before spawning the next one.

8. Idle Detection

When an agent goes idle without a completion report:

If --no-auto-resume flag is set: Immediately escalate to user. Do not attempt auto-recovery.

Otherwise:

Classify using observable signals:

• NO_REPORT — no TaskUpdate and no messages from the agent. → Resume with: "Execute the test scenario now. Take a snapshot first to see the current page state."

• STALLED — agent sent an error message but went idle without resolving. → Resume with the quoted error + "Try an alternative approach or report the failure."

Progress signals: Agents send SendMessage after each major step (setup complete, each step executed, asserting outcomes). Use these to classify idle state.

Escalation: After 2 failed resume attempts (no progress between attempts), escalate to user: "Test scenario T{N} agent is stuck after 2 resume attempts. Last known state: {agent's last message or 'no messages'}. How to proceed?"

9. Update Living State Doc

After all scenarios complete (or are skipped), append a new ### Run: entry to the ## Test Results section. Create the section if it doesn't exist.

Format:

### Run: {YYYY-MM-DD HH:MM}
**Base URL:** {url}
**Execution:** serial (1 agent at a time)
**Overall:** {X passed, Y failed, Z skipped} / {total}

| # | Scenario | Result | Notes |
|---|----------|--------|-------|
| T1 | {title} | Pass | — |
| T2 | {title} | Fail | {brief reason} |
| T3 | {title} | Skipped | dependency T2 failed |

### Failure Details
#### T2 — {title}
- **Step failed:** {step description}
- **Expected:** {what was expected}
- **Actual:** {what happened}
- **Screenshot:** {path or "N/A"}

Previous run entries are preserved — always append, never overwrite.

10. Cleanup

1. TeamDelete("test-{slug}")
2. Delete .project-orchestrator/state.json
3. No scope file to clean up (test agents are read-only)

11. Present Summary

Format:

Test Results: {X passed, Y failed, Z skipped} / {total}

{If failures, list each with scenario title + brief reason}
{If screenshots captured, list paths}

Next steps based on results:

• All passed: "All tests passed. Next: /project:verify then /project:finish"
• Some failed: "{X} tests failed. Review failures above, fix issues, and re-run /project:test"
• All skipped: "All tests were skipped due to dependency failures. Fix the root failure and re-run."