implement-tasks
Implement Tasks
Turn an approved queue into a verified change. Be a thin conductor: own model selection, scheduling, integration, and delivery. implement-and-review owns each task's implementation, focused review, and acceptance evidence.
The sequence is interview-me → to-prd → to-tasks → implement-tasks. Receive the approved PRD and task queue under .ai-workflow/work/<feature-slug>/. A complete single task uses the same model approval and skips graph scheduling. Missing product decisions return to the appropriate earlier stage.
Select the current phase
| Phase | Required work and reference |
|---|---|
| Prepare or revise a model plan | Read references/model-plan.md. Validate the approved queue, dependencies, input hashes, routes, and available transports; produce the exact preview. |
| Schedule and integrate approved tasks | Read references/feature-orchestration.md. Dispatch ready tasks, enforce write ownership, and verify the combined result before releasing dependents. |
| Verify and deliver | Read references/completion.md. Reconcile cross-task risks, capture final acceptance, and deliver within the user's authorization. |
| Resume | Compare saved scope, route approval, task inputs, and repository state before continuing the applicable phase. Reuse matching approval and evidence. |
Resolve shared/ as the collection's shared skill describes and worker skills from their loaded locations. Use shared/references/workflow-stage-routing.md for engineering placement, shared/references/task-shaped-model-routing.md and shared/references/model-roster.md for model selection, and the shared routing schema and run-state contract for persisted state. Load these at the phase that uses them and reuse active context.