Dispatch Programmers

Step 1: Load Plan and Establish Recovery State

1. Load plan via get_plan (brain-mcp tool) using the plan:// URL. If given a file path instead, read the file — but plans should be stored in the graph.
2. Extract ALL tasks with full text and context
3. Create TodoWrite with all tasks
4. Note any cross-task dependencies
5. Record base SHA: git rev-parse HEAD — this is the recovery checkpoint
6. Check for prior progress: If resuming after a session interruption, check TodoWrite state and git log to identify which tasks already completed. Skip completed tasks.

Recovery detection: If the plan exists (via get_plan) and TodoWrite has tasks marked complete, this is a resumed session. Verify completed tasks by checking git log for their commits, then continue from the first incomplete task. Update task statuses in the plan via update_plan as work progresses.

Step 2: Per-Task Cycle

Dispatch implementer subagent with isolation: "worktree":

• Full task text from plan (don't make subagent read plan file)
• Scene-setting context (where this task fits in the project)
• Clear goal and constraints
• Skill loading: subagent MUST load /implement + the appropriate language skill from the dispatch table in /programming
• Explicit commit requirement: subagent MUST commit before returning

Implement Task N: [name]

Context: [brief description of where this fits]

[Full task text from plan, verbatim]

Requirements:
- Load /implement and follow its TDD loop
- Load [language skill] for patterns and conventions
- The /implement loop will dispatch /audit-code after each round
- Loop continues until the auditor returns a clean report
- Commit when audit passes with message: "feat(task-N): [description]"
- Write implementation decisions to statement-mcp as handoff messages
- Return summary of what you built, audit rounds needed, and the commit SHA

Worktree isolation: Each implementer runs in its own worktree. This means:

• Partial work from a crashed subagent doesn't pollute the feature branch
• The worktree can be discarded if the subagent fails
• Successful work is merged back explicitly

After subagent returns: Merge its worktree branch back to the feature branch. This is the durable checkpoint — if the session dies after this merge, the task's work is preserved.

If implementer asks questions: Answer clearly and completely. Provide additional context. Don't rush them into implementation.

After merge: Mark task complete in TodoWrite immediately. The /implement loop already ran /audit-code until clean — no additional review dispatch needed per task.

Session Recovery

If the session is interrupted at any point, the next session can recover:

Key invariant: The feature branch only receives complete, tested, committed work. Partial work lives in worktrees that can be safely discarded.

Step 3: Branch Completion

After all tasks complete:

1. Cross-task integration audit — dispatch /audit-code across the ENTIRE implementation (all tasks combined). This catches interaction effects between independently-implemented tasks that per-task audits couldn't see.
2. If audit finds issues — dispatch fix subagent (worktree-isolated) → fix → re-audit until clean
3. Verify tests pass — run full suite, read output, confirm zero failures
4. If tests fail — fix before offering options, do not proceed

Present exactly these options:

Implementation complete. What would you like to do?

1. Merge back to <base-branch> locally
2. Push and create a Pull Request
3. Keep the branch as-is (I'll handle it later)
4. Discard this work

Option 1 — Merge locally: checkout base, pull latest, merge feature, verify tests on merged result, delete feature branch.

Option 2 — Create PR: push branch, gh pr create with summary and test plan.

Option 3 — Keep as-is: report branch name and path, preserve everything.

Option 4 — Discard: require typed "discard" confirmation. Show what will be deleted (branch, commits, worktree). Only then delete.

Handling Review Feedback

When receiving review feedback from any source:

• Verify before implementing — check suggestion against codebase reality
• Push back if wrong — technical correctness over social comfort. If suggestion breaks things, lacks context, or violates YAGNI, say so with reasoning.
• Clarify all items first — if any feedback is unclear, ask before implementing anything. Partial understanding = wrong implementation.
• One fix at a time, test each — don't batch fixes without testing between them
• No performative agreement — don't say "Great point!" or "You're absolutely right!" Just fix it or push back.

Red Flags

Never:

• Start implementation on main/master without explicit user consent
• Skip the /implement audit loop (every task gets audited until clean)
• Skip the cross-task integration audit after all tasks complete
• Dispatch multiple implementation subagents in parallel (conflicts)
• Make subagent read plan file (provide full text instead)
• Proceed with failing tests
• Trust agent success reports without independent verification
• Force-push without explicit request
• Leave a task unmarked in TodoWrite after its work is merged
• Let a subagent return without committing (uncommitted work = unrecoverable work)
• Let a subagent skip loading /implement or the language skill

If subagent asks questions: Answer before letting them proceed.

If subagent fails: Dispatch fix subagent with specific instructions. Don't fix manually (context pollution). Use worktree isolation.

If cross-task audit finds issues: Dispatch fix subagent (worktree-isolated) → fix subagent commits → merge back → re-audit. Don't skip re-audit.

If session may be interrupted: Every commit on the feature branch is a recovery checkpoint. TodoWrite state plus git log are sufficient to resume from any point.

When NOT to Use

• Tasks are tightly coupled and can't be understood independently
• No written plan exists (use brainstorming → writing-plans first to create a plan:// node)
• Single small task (just do it directly)
• Exploratory debugging (use systematic-debugging instead)

Real-World Impact

From production sessions: 18 tasks across 3 projects (xnn, OnShape, focus_stacking) executed via this workflow. Spec compliance review caught over-building (extra flags not in spec) and under-building (missing progress reporting). Code quality review caught magic numbers and duplicate code. Two-stage review found issues that single-pass review missed — spec compliance ensures correctness, quality review ensures craftsmanship.