implement-tasks
Implement Tasks
Turn an approved queue into one verified change. Be a thin conductor: you own the model plan, scheduling, write ownership, dependency release, integration, and delivery. implement-and-review owns everything inside a task: its roles and contexts, review cycles, isolation, the implementer's brief, and evidence. Do not restate or override those rules here.
The input is the approved PRD and tasks.md from to-tasks under .ai-workflow/work/<feature>/, laid out as shared/references/work-state.md describes. A bare plan goes to to-prd and to-tasks first. One bounded task with a complete acceptance contract and no open product, security, contract, data, or dependency decision goes straight to implement-and-review.
Plan
- Read
tasks.mdand each task'sbrief.md. An open decision (a task markedDesign: openis not blocked; its debate settles the design), missing access, a dependency cycle, or two tasks owning the same path without a merge plan blocks the affected tasks. - Pick the review mode. A new feature queue uses feature mode: mark tasks
criticalwhen they carry high risk or a shared contract, and defer the rest to one final review. Otherwise use task mode, where every task gets its own reviewer. An existing approved plan keeps its mode and routes. - Resolve each task's route with
python3 shared/scripts/model_routing.py resolve <route> --provider <host> [--risk high], giving--risk highonly for a task that touches the invariantspolicy.high_risk_triggersnames, never for a sensitive feature label. For a task whose brief saysDesign: open, add the debate seats (challenger and arbiter) from thedesign-debateroute to its row. Add the integration or final reviewer, whose model differs from every author's. Read.rar-skills/config.local.yamlonce, if present, as advisory input while forming the plan (shared/references/local-config.md); never reread it after approval. - Show the plan: for each task its roles with model, effort, and host route; the integration or final reviewer; the concurrency cap; the review-cycle limit; and the total call ceiling. Ask: "Approve this plan, or specify changes." A request to use defaults and run approves the unchanged resolved plan. Task approval,
--auto, a saved preference, or silence is not model approval. When the user wants one decision for queue and models, show both together; the single response approves both. - Write the approved plan to the top of
tasks.mdwith the user's response and the time. It is fixed from then on. A change to any model, effort, role, host route, or limit needs the affected rows approved again. An unavailable route blocks its own tasks; never swap in another model.
Run
Keep the ledger, not the transcript: tasks.md plus the task reports are the ledger. Update a task's row before dispatch and after its result, and resume from them.
Each wave: