AI Dispatch

Dispatch

Purpose

Execution engine for approved plans. Reads plan.md and tasks.md, dispatches one subagent per task (fresh context), runs two-stage review on each deliverable, and tracks progress. If stuck: STOP and re-plan.

When to Use

• After /ai-plan produces an approved plan
• To resume execution: /ai-dispatch --resume
• Never without an approved plan (run /ai-plan first)

Process

1. Load plan -- read specs/spec.md -> specs/plan.md
2. Load decisions -- read decision-store.json for constraints
3. Build DAG -- parse task dependencies, identify parallel groups
4. Execute phase by phase -- for each phase: a. Dispatch one subagent per task (fresh context window) b. Each subagent receives: task description, file scope, boundaries, constraints c. Run two-stage review on deliverable (see below) d. Update task status in tasks.md e. Check phase gate before advancing
5. Track progress -- update tasks.md checkboxes after each task
6. Report completion -- summary of all tasks, any concerns raised

Task Statuses

Two-Stage Review

Every task deliverable goes through two reviews before marking DONE:

Stage 1: Spec Compliance

• Does the deliverable match the task description?
• Does it satisfy the acceptance criteria from spec.md?
• Are all file scope boundaries respected (no out-of-scope changes)?

Stage 2: Code Quality

• Stack validation passes (ruff, tsc, cargo check, etc.)
• No new lint warnings introduced
• Test coverage maintained or improved
• No governance advisory warnings from guard

If either stage fails: fix and re-review (max 2 retries per stage).

DAG Construction

Independent (can run in parallel):

• Different file scopes with no overlap
• No producer-consumer relationship
• Different modules with no shared state

Dependent (must serialize):

• Task B reads files Task A creates
• Task B depends on Task A's output
• Both modify governance artifacts (.ai-engineering/ must serialize)
• Plan explicitly orders them

Subagent Context

Each subagent receives a focused context window:

task: T-2.1
description: "Implement the parse_config function"
agent: build
scope:
  files: ["src/config.py", "tests/test_config.py"]
  boundaries: ["Do NOT modify src/main.py", "Do NOT touch hooks/"]
constraints:
  - "Follow existing ConfigParser pattern in src/base_config.py"
  - "TDD: test files from T-2.0 are IMMUTABLE"
gate:
  post: ["ruff check", "pytest tests/test_config.py"]

Stuck Protocol

If a task fails after 2 retries:

1. Mark task as BLOCKED with reason
2. Check if other tasks in the phase can proceed independently
3. If phase is blocked entirely: STOP execution
4. Report to user: what failed, what was tried, options (re-plan, skip, manual fix)

Never loop silently. Never retry the same approach more than twice.

Progress Tracking

Update tasks.md in real-time:

- [x] T-1.1: Create config module @build -- DONE
- [x] T-1.2: Add validation logic @build -- DONE_WITH_CONCERNS (perf warning)
- [ ] T-2.1: Write integration tests @build -- IN PROGRESS
- [ ] T-2.2: Security scan @verify -- PENDING

Common Mistakes

• Dispatching without an approved plan.
• Giving subagents the entire codebase context (scope them tightly).
• Skipping the two-stage review.
• Continuing past a BLOCKED task without user input.
• Modifying test files from a RED phase during a GREEN phase task.

Integration

• Called by: user directly (after /ai-plan approval)
• Calls: ai-build agent (build tasks), /ai-verify (scan tasks), /ai-guard (gate checks)
• Transitions to: /ai-commit (after all tasks DONE), or back to /ai-plan (if re-plan needed)

$ARGUMENTS