orchestrate

name	orchestrate
description	Orchestrates execution of a plan by spawning agents to work on tasks according to the plan's sequencing instructions. Reads a plan from context/plans/ready/, moves it to in-progress, and coordinates task execution across phases. Use when a reviewed plan is ready for implementation.
disable-model-invocation	true
argument-hint	[plan-name]

Orchestrate

Orchestrate a plan from context/plans/ready/. Coordinate — don't produce. Dispatch agents, track progress.

Available plans

!ls context/plans/ready/ 2>/dev/null || echo "(none)"

Process

1. Load the plan

Read these context library files first:

• context/lib/index.md — agent router, architectural principles
• context/lib/development_guide.md — conventions, constraints, coding standards
• context/lib/testing_guide.md — what to test, test patterns

Then read context/plans/ready/$ARGUMENTS/index.md. If missing, list available plans and ask which to run.

Understand:

• Shared context section (every agent needs this)
• Each task's description and acceptance criteria
• Sequencing: phases, concurrency, and dependencies

2. Move to in-progress

git mv context/plans/ready/<plan-name> context/plans/in-progress/<plan-name>

Commit the move.

3. Execute phases in order

For each phase in the sequencing section:

Sequential: One worker agent at a time. Use model: "gpt-5.5" with reasoning_effort: "high" for complex, cross-cutting, architectural, or integration-heavy tasks, and reasoning_effort: "medium" for bounded implementation tasks. Wait for completion before starting the next.

Concurrent: Spawn all independent phase worker agents simultaneously via multiple spawn_agent calls in one message. Use model: "gpt-5.5" and choose reasoning_effort: "medium" or "high" per task complexity. For each agent, provide:

1. The plan's Shared Context section
2. The agent's specific task — description, acceptance criteria
3. Instruction to read relevant context/lib/ files for architectural guidance
4. Instruction to follow context/lib/development_guide.md conventions
5. Instruction to run cargo check and cargo test before considering the task complete

Do NOT provide:

• Other tasks' details (the agent doesn't need them)
• The full plan document (wastes context)
• Freedom to expand scope beyond acceptance criteria

4. Integrate results

After each phase:

• Review what agents produced
• Verify acceptance criteria are met
• If a task completed partially or blocked, surface to the user with context
• If using worktrees, merge completed work back to the main branch

Between phases, check that prerequisites for the next phase are satisfied.

5. Complete

When all phases are done:

• Run preflight checks: cargo fmt --check && cargo clippy -- -D warnings && cargo test
• Run a /review-panel on code edited in this session
• Report review panel findings to user to discuss which feedback to act on

6. Landing the plane

When the user says "land the plane":

• Move the plan to done: git mv context/plans/in-progress/<plan-name> context/plans/done/<plan-name>
• Clean up worktrees from the session
• Commit & push

Error handling

• Agent fails a task: Surface the error and acceptance criteria to the user. Ask whether to retry, skip, or abort.
• Merge conflict from concurrent agents: Resolve if straightforward; escalate to user if the conflict involves architectural decisions.
• Preflight fails: Fix if the issue is mechanical (formatting, simple clippy lint). Escalate if the fix requires design decisions.

Principles

• You coordinate, you don't produce. Every tool call spent building is context not spent orchestrating.
• Guard context. Each agent gets minimum viable context for their task.
• 3 of 4 completing is enough. Partial progress with clear status beats blocking on one stuck task.
• Surface, don't guess. Tell the user when something unexpected happens. Don't make architectural decisions on their behalf.