execute

Execute — Execute Phase

Load an intent-driven plan, group tasks into dependency-ordered waves, dispatch specialist agents in parallel within each wave, run language-aware verification and a mandatory critic gate between waves, then dispatch the sentinel for final sign-off.

This gate is mechanically enforced, not just instructed: while a run is armed (see Step 1), the plugin's PreToolUse hook denies dispatch of any agent other than critic or welder once a code-writing agent has completed and the critic has not yet reviewed. Arm the gate at run start and disarm it on every exit path.

Process Flow

Start
  |
  v
[1. Load plan]
  Parse tasks, dependencies, agent assignments, verify commands.
  |
  v
[2. Detect agents]
  Is fakoli-crew installed? Log available agents. Note fallback.
  |
  v
[3. Group into waves]
  Apply dependency rules (see Wave Assignment section below).
  |
  v
[For each wave:]
  |
  v
  [a. Dispatch agents in parallel]
     One Agent tool call per task. All tasks in the wave start simultaneously.
     |
  [b. Wait for completion]
     Read agent status files. Poll until all show COMPLETE, BLOCKED, or NEEDS_REVIEW.
     |
  [c. Handle escalations]
     BLOCKED -> surface to user, wait for resolution, then re-dispatch.
     NEEDS_REVIEW -> surface to user, wait for decision, then continue.
     |
  [d. Run language verification]
     TypeScript: npx tsc --noEmit
     Python:     ruff check . && mypy .
     Rust:       cargo check
     If fail -> dispatch welder to fix -> re-run verification.
     Do NOT proceed to critic until verification passes.
     |
  [e. Run critic gate (non-negotiable)]
     Collect modified files from status files.
     Dispatch critic with file list + acceptance criteria.
     PASS or SHOULD FIX/NIT -> proceed.
     MUST FIX -> enter fix cycle (max 3 iterations, then escalate).
  |
  v
[After all waves: Dispatch sentinel]
  Final verification with evidence gate.
  |
  v
[Report summary]