execution-coordination

Files (4)

• SKILL.md
• delegation-prompting.md · resources/
• integration-review.md · resources/
• parallel-seam-selection.md · resources/

SKILL.md (first lines)

# Execution Coordination

This is an internal meta skill. Humans should not need to reference it directly.

## Purpose

Give Darwin a dedicated internal home for subagent delegation, execution sequencing, review checkpoints, and final integration.

The goal is to keep orchestration logic out of `build` while making parallel and delegated execution systematic instead of ad hoc.

## Resources

- `resources/parallel-seam-selection.md` for deciding whether to split work
- `resources/delegation-prompting.md` for packaging subagent context
- `resources/integration-review.md` for reviewing and integrating returned work

## When To Use

Use when:

- there is an approved implementation plan
- work can be split into independent seams or tasks
- a task benefits from isolated context in a fresh agent
- review and reintegration need to be explicit

## Core Rule

Do not delegate blindly.

Only dispatch work when the task boundary, required context, and verification expectation are clear enough for a subagent to succeed without inheriting the full orchestrator session.

## Workflow

1. Read the approved plan and identify independent seams or tasks.
2. Use `parallel-seam-selection` to decide which tasks can run in parallel and which must stay sequential.
3. Use `delegation-prompting` to package the minimum necessary context for each subagent.
4. Dispatch fresh subagents for implementation or investigation work.
5. Use `integration-review` to review the returned work for spec compliance and integration fit.
6. Integrate or request fixes before moving the plan forward.
7. Keep plan state accurate as tasks move from pending to done.

## Delegation Rules

- Prefer parallel agents for independent seams like data contracts, clients, mapping, state, and UI integration.
- If there are 2 or more independent seams with mostly separate files, parallel subagents are the default.
- Prefer sequential execution when later work depends directly on earlier edits.
- Keep the orchestrator focused on task shaping, review, and integration.
- Do not make subagents rediscover plan context that can be handed to them precisely.

## Review Rules

- Check that returned work matches the assigned task, not just that code exists.
- Verify the promised test or check was actually run.
- If the subagent flagged uncertainty, address it before treating the task as complete.
- If the output does not fit the current branch state or plan, send it back for revision or integrate carefully with explicit reasoning.

## Guardrails

- Do not dispatch multiple agents against the same files unless coordination is explicit.
- Do not hoard all implementation in the orchestrator when seams are independent.
- Do not let a subagent's success claim replace verification.
- Do not mark a plan step complete until the work has been reviewed and integrated.

## Output

When using this skill, summarize execution in terms of:

- what was delegated
- what ran in parallel versus sequentially
- what reviews or corrections were needed
- what is now integrated and verified

From tykowale/darwin, MIT.

More from tykowale/darwin