Orchestrate — Complex Task Orchestration

Apply workflow orchestration methodology for multi-step tasks with 3+ steps or architectural decisions.

How to use

• /orchestrate <task> — plan-first approach for any non-trivial task

---

Methodology

1. Plan first (invoke writing-plans skill)

Always start by invoking the writing-plans skill (or native plan mode if your platform has it) to produce a structured plan with TDD-style task steps. Write it to wherever writing-plans puts plans for this repo — do not assume a fixed path. Do not proceed to coding until the plan is written and reviewed.

• If something goes sideways mid-task: STOP, re-plan using writing-plans again
• Track progress via tasks/todo.md checkboxes

2. Use subagents (fan out independent work)

• Offload research, exploration, and parallel analysis to subagents
• Keep main context window clean and focused
• One task per subagent for precision
• Invoke the subagents skill for the fan-out ladder and standard return contract

Once the plan splits into independent workstreams, do not implement them serially. Fan them out using the ladder from the subagents skill (Workflow tool → parallel subagents → serial fallback):

plan (writing-plans)
  → parallel implementation subagents     # one per independent workstream, isolated
      each returns { summary, files[], diff, risks[], confidence }
  → gather the structured diffs
  → independent reviewer subagent          # fresh context, runs /code-review — no self-review bias
      returns { findings[]{severity,file,detail}, verdict }
  → merge / verify gate                    # apply, run /gate, only then mark done

Reference dispatching-parallel-agents and subagent-driven-development for the mechanics. Use worktree isolation when subagents edit files in parallel so they don't collide. Workstreams that share state (same module, ordering dependency) stay serial.

3. Self-improvement loop

• After any correction: note the pattern in tasks/lessons.md
• Review lessons at session start for relevant context

4. Verify before done

• Never mark complete without proving it works
• Run tests, check logs, diff behaviour vs main
• Ask: "Would a staff engineer approve this?"

5. Demand elegance

• After a working but messy fix: "Implement the elegant solution"
• Skip this for simple, obvious fixes

6. Autonomous bug fixing

• When given a bug: investigate and fix. Do not wait for hand-holding.
• Pair with /fix for CI failures and log-based bugs.

---

Task Management

1. Write plan → tasks/todo.md
2. Check in before implementation starts
3. Mark items complete as you go
4. Document results in review section
5. Capture lessons → tasks/lessons.md