Delegation Autopilot

When to Use

Use native subagents before direct edits when at least one is true:

• Work can be split into independent tracks.
• Request spans multiple files/modules.
• Request includes implementation plus tests/docs/review.
• User explicitly asks for delegation/subagents.

Do not use for tiny single-file edits, strict sequential chains, or pure explanations.

Deterministic Orchestration Rubric

1. Tiny/sequential/local task -> no delegation.
2. Single non-trivial track -> one spawn_agent.
3. Independent tracks A/B/C -> one spawn_agent per track, then grouped wait.
4. Strict track -> one dedicated worker with explicit constraints and parent checkpoints.

Exception: when explicitly following /Users/kosta/.codex/superpowers/skills/subagent-driven-development/SKILL.md, keep implementation subagents sequential for that workflow's staged review loop.

Default Native Profile

Role defaults:

• Research/review/audit -> agent_type: "explorer"
• Implementation -> agent_type: "worker"
• Final verifier -> agent_type: "explorer" (or default)

Parallel defaults:

• Spawn all independent track agents up front.
• Wait on all IDs together.
• Keep a parent ledger of track -> agent_id -> status.
• Treat agent_id as the durable persistence key (replacing old MCP run_dir references) for resume, verifier context, and recovery.

Mandatory Constraints in Every Subagent Prompt

• Do not spawn subagents.
• Do not call MCP delegate_* tools.
• Keep scope bounded to assigned files and outcomes.
• Return the required report contract exactly.

Mandatory Report Contract (native)

Every subagent must report:

1. agent_id
2. Files read
3. Files changed (if any) + rationale
4. Tests/checks run + pass/fail
5. Blockers/open questions
6. Risks/integration notes
7. Recommendations/next actions

Final Verifier Loop (required)

After implementation tracks complete:

1. Spawn one final verifier subagent.
2. Provide changed-file list + track summaries.
3. Verify cross-track consistency and regression risk.
4. Integrate only after verifier report is reviewed.

Timeout and Recovery

If one or more agents are incomplete:

1. Re-run wait with explicit IDs and higher timeout.
2. Send corrective guidance via send_input.
3. Use resume_agent for interrupted threads.
4. close_agent completed tracks to keep orchestration clean.

Long-Session Anti-Drift Rule

Every 3-5 turns in long delegated sessions, restate:

• no subagent recursion
• parent-orchestrated one-agent-per-track model
• grouped wait for true parallel completion
• mandatory final verifier gate

MCP -> Native Translation Table

• delegate_autopilot(task, ...) -> split task into tracks + spawn_agent per track + grouped wait
• delegate_run(...) -> strict single-track spawn_agent with deterministic prompt
• delegate_resume(thread_id, ...) -> resume_agent(id)
• MCP aggregate/run metadata -> parent-maintained ledger keyed by agent_id
• max_agents/max_parallel knobs -> explicit track decomposition + spawn count

Parent Orchestrator Pseudocode

1) Build tracks: T1, T2, T3 (independent)
2) id1 = spawn_agent({agent_type:"worker", message:"<T1 prompt>"})
3) id2 = spawn_agent({agent_type:"worker", message:"<T2 prompt>"})
4) id3 = spawn_agent({agent_type:"explorer", message:"<T3 prompt>"})
5) wait({ids:[id1,id2,id3], timeout_ms:120000})
6) If needed, send_input({id:id2, interrupt:false, message:"<follow-up>"})
7) wait({ids:[id2], timeout_ms:120000})
8) verifier = spawn_agent({agent_type:"explorer", message:"<final verifier prompt>"})
9) wait({ids:[verifier], timeout_ms:120000})
10) close_agent for all completed agents

Prompt Template: Implementation Track

## Task
<concise track task>

<scope>
- In scope: <files/areas>
- Out of scope: <explicit exclusions>
</scope>

<constraints>
- Do not spawn subagents.
- Do not call MCP `delegate_*` tools.
- Keep edits minimal and reversible.
- Run appropriate tests/checks; if skipped, explain why.
</constraints>

<deliverable>
1. agent_id
2. Files read
3. Files changed + why
4. Tests/checks + results
5. Blockers/open questions
6. Risks/integration notes
7. Next actions
</deliverable>

Prompt Template: Final Verifier

## Task
Final integration verification for all completed tracks.

<inputs>
- Changed files by track
- Track outputs and assumptions
</inputs>

<checks>
1. Cross-file naming/reference consistency
2. Import/dependency/interface compatibility
3. Test alignment with touched behavior
4. Obvious regressions/conflicts
</checks>

<constraints>
- Do not spawn subagents.
- Do not call MCP `delegate_*` tools.
- Read-only unless explicitly asked to patch.
</constraints>

<report>
1. agent_id
2. Files reviewed
3. Findings by severity
4. Suggested minimal fixes
5. Verification commands
</report>

Safety

• Never expose secrets from local config.
• Keep prompts scoped and explicit.
• Parent agent retains final integration ownership.