Task Run

Mode: Work on all tasks or a specific one

⚠️ STRICT WORKFLOW - NO IMPROVISATION

You MUST follow this exact sequence for EVERY task. Do NOT skip or reorder steps.

next → implement → check → done

⛔ CLI-ONLY ENFORCEMENT

NEVER bypass CLI for workflow decisions:

This applies to ALL iterations in the loop.

⚡ UNATTENDED MODE (when no task_id provided)

• NO questions allowed
• NO stopping for errors
• MUST complete all tasks

Mode Detection

If task_id provided (e.g., feature-run auth.login):

• Work on that specific task only
• Complete it and stop

If no task_id (e.g., feature-run):

• Auto-complete all pending tasks
• Loop until all done
• UNATTENDED MODE ACTIVE - see rules below

Single Task Mode

When task_id is provided:

# STEP 1: Delegate to implementer agent
Task(
  subagent_type="agent-foreman:implementer",
  prompt="TASK: <task_id>
Execute: next <task_id> → implement → check → return result"
)

# STEP 2: Parse result and complete
# - If success + verification_passed: agent-foreman done <task_id>
# - Otherwise: agent-foreman fail <task_id> -r "<notes>"

All Tasks Mode

When no task_id:

# STEP 1: Show initial state (once only)
agent-foreman status

# STEP 2: LOOP
  # 2a: Delegate to implementer agent (auto-select task)
  Task(
    subagent_type="agent-foreman:implementer",
    prompt="Execute: next → implement → check → return result"
  )

  # 2b: Parse ---IMPLEMENTATION RESULT--- from agent output
  # Extract: task_id, status, verification_passed, notes

  # 2c: Check if implementer returned "No pending tasks"
  #     If yes → EXIT to STEP 3
  #     If no → Continue to 2d

  # 2d: Handle result
  # - success + verification_passed → agent-foreman done <task_id>
  # - Otherwise → agent-foreman fail <task_id> -r "<notes>"

  # 2e: Continue loop (go to step 2a, NOT step 1)

# STEP 3: FINISH
# Show final summary with agent-foreman status (once only)

Rules

Priority Order

1. needs_review → may be broken (highest)
2. failing → not implemented
3. Lower priority number → higher priority (0 is highest)

TDD Mode

TDD workflow is handled internally by the agent-foreman:implementer agent. No action required from the orchestrator.

Unattended Mode Rules

When in ALL-TASKS mode (no task_id provided):

• NO questions - The implementer agent handles decisions autonomously
• NO stopping - Always continue to next task after done/fail
• Exit only when implementer returns "No pending tasks available"

Loop Enforcement (MUST FOLLOW)

When in ALL-TASKS mode:

1. START: Run agent-foreman status once - show initial state
2. LOOP:

a. Delegate: Call Task(subagenttype="agent-foreman:implementer") - agent runs next → implement → check b. Parse result: Extract taskid, status, verificationpassed, notes c. Check exit: If status: blocked and notes contains "No pending tasks" → EXIT to step 3 d. Complete: Based on result: - success + verificationpassed: true → agent-foreman done <taskid> - Otherwise → agent-foreman fail <taskid> -r "<notes>" e. Continue: Go to step 2a (NOT step 1)

1. FINISH: Run agent-foreman status once - show final summary

Result Parsing

The implementer agent returns structured output:

---IMPLEMENTATION RESULT---
task_id: <task_id>
status: success|partial|blocked|failed
verification_passed: true|false
files_modified: [list]
notes: <description>
---END IMPLEMENTATION RESULT---

Never Stop For

• Agent errors
• Verification failures
• Missing tasks
• Any unexpected output

Only Stop When

• Implementer agent returns status: blocked with notes: No pending tasks available

On Agent Result Handling

When the implementer agent returns its result:

If status: success and verification_passed: true

agent-foreman done <task_id>

Then continue to next iteration.

If verification failed or status is not success

agent-foreman fail <task_id> --reason "<notes from agent>"

Then IMMEDIATELY continue to next iteration.

Result Handling Rules

1. DO NOT STOP - This is the most critical rule
2. DO NOT ASK - Never ask user what to do
3. ALWAYS parse the result - Extract task_id and status from agent output
4. ALWAYS continue - Move to next task after done/fail

This applies to ALL agent results, including errors or malformed output.

Exit When

CRITICAL: NEVER stop due to verification failure - always use agent-foreman fail and continue!

Loop Completion

When all tasks have been processed:

1. Run agent-foreman status to show final summary
2. Report counts:

- X tasks passing - Y tasks failed (need investigation) - Z tasks needs_review (dependency changes) - W tasks still failing (not attempted)

1. List tasks that failed verification with their failure reasons