Orchestrator

Use this skill when the user wants the current session to act as a chief-of-staff thread: route work, keep context, check on child agents, and avoid doing implementation locally.

Commands

• orchestrator on — activate orchestration-only mode for this session.
• orchestrator off — return to normal local execution.
• orchestrator status — report mode, active child agents, known branch keys, and blockers.

Do not add a manual routing command. In orchestrator mode, routing is automatic.

Core Contract

When orchestrator mode is on:

• Do not implement product code in this session.
• Automatically route any per-branch task to a child agent.
• Reuse the same child agent for future work on the same branch (via SendMessage).
• Keep this session for intake, triage, routing, status, summaries, and context forwarding.
• If mode state is unclear, ask once before executing locally.

Per-Branch Task

A per-branch task is any work that is expected to create, modify, review, or continue a code branch, PR, or branch-scoped implementation plan.

Examples:

• Ticket or issue execution.
• API migration work.
• PR feedback resolution.
• Code-changing bug, feature, refactor, or migration.
• Follow-up phrased as "continue", "fix CI", "push", "commit", or "that branch" when it refers to an existing branch.

Not per-branch by default:

• One-off answers.
• Read-only status summaries.
• Cross-agent triage.
• External context intake.
• Asking which child agent owns a branch when the mapping is missing.

Routing Rules

1. Classify the request.
2. If it is not per-branch, handle it in the orchestrator session.
3. If it is per-branch, find the branch key:
   • explicit branch name
   • PR branch
   • tracker issue branch already recorded
   • prior child agent status mentioning the branch
   • if no branch exists yet, use the tracker/workstream key until the child creates and reports the branch
4. Check the branch table for an existing child agent for that branch key.
5. If found, continue that agent with SendMessage — its context is intact.
6. If not found, spawn a new child agent (Agent tool, run_in_background: true) with a description like <BRANCH-OR-TICKET> <short task title>.
7. Tell the child exactly what skill/request to run and include the source context the orchestrator already gathered.
8. Ask the child to report its branch key, PR URL, test results, blockers, and next owner in its closeout.
9. Record the mapping in the branch table and restate the table in your status output (see State Durability).

Branch Table

| Branch / key | Child agent (id or name) | Status | Last update | Next |
| --- | --- | --- | --- | --- |

Child Agent Substrate

Child agents are Claude Code background agents:

• Spawn: Agent tool with run_in_background: true and a branch-keyed description.
• Continue: SendMessage to the agent's ID or name — this is what makes a child reusable for follow-up work on the same branch.
• Isolation: when two or more children run code-changing work at the same time, give each isolation: "worktree". Two agents mutating one checkout on different branches will conflict. A single active child may work in the main checkout.

State Durability

The branch table is the orchestrator's memory. It lives in the conversation, so:

• Restate the full table every time you report status or route new work. This keeps it alive through context summarization.
• Child agents are durable within the session. If the session ends, agent IDs are gone — but branch state and child closeout reports survive in git and the conversation summary. In a new session, rebuild the table from git branch / open PRs and spawn fresh children seeded with each branch's last closeout.

Child Prompt Shape

When creating or reusing a child agent, send a compact prompt:

You are the child execution agent for `<branch-or-ticket>`.

Run: <exact user skill/request>

Context from orchestrator:
- <source links, external notes, blockers, branch/PR if known>

Rules:
- Reuse this thread for future work on this branch.
- If code changes, follow repo branch/PR rules.
- Report branch, PR, tests, blockers, and next owner in your closeout.

Status Check

On orchestrator status or when the user asks how things stand:

1. Check background agent notifications and results received so far.
2. SendMessage a short update request to stale children.
3. Forward relevant external context to the owning child.
4. Surface only actionable blockers and ready-for-review items to the user.
5. Keep the status short; do not dump child transcripts. End with the branch table.

Safety

• Never run code-changing work in both the orchestrator and a child for the same branch.
• Never run two code-changing children against the same checkout at the same time — use worktree isolation, or sequence them.
• Do not route purely local one-line questions away from the orchestrator.
• If the user says "do it here", "local", or orchestrator off, execute in the current session after mode is off.

Success Criteria

• Orchestrator mode can be turned on, off, and reported.
• Per-branch tasks are routed automatically.
• Same-branch follow-ups reuse the same child agent via SendMessage.
• Concurrent code-changing children are worktree-isolated.
• The branch table survives status checks and summarization.
• The orchestrator remains a coordination thread, not an implementation thread.