mcpField
I
implement-and-review

implement-and-review

github.com/robsonrungDevelopment266 installsUpdated 2d ago
Repository

About implement-and-review

An agent skill for implementing an approved task or completing a bound feature review against its acceptance contract. It defines work paths, execution and review context handling, evidence, and who can accept the result.

Key features

  1. 01Selects task or review path by policy
  2. 02Sets rules for routes and context handling
  3. 03Defines verification evidence and report fields
  4. 04Separates worker results from final acceptance

Use cases

  1. 01Implement an approved task with clear scope
  2. 02Complete a combined feature review under policy
  3. 03Repair and recheck work on the same path

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/robsonrung/rar-skills --skill implement-and-review

After installing, type/implement-and-reviewin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

agents/
openai.yaml
references/
deferred-task.md
evidence-strategy.md
final-feature-review.md
methodology.md
review-brief.md
runner-invocations.md
task-implementation.md
task-review.md
worktree-and-integration.md
scripts/
launch.py
tests/
test_launch.py

Details

From the project README

Implement And Review

Build one approved task to its acceptance contract, or complete its bound feature review. A worker result is integration evidence. The parent alone accepts tasks after the applicable gate passes.

Authority and common rules

An implementation request authorizes in-scope edits, stated verification, and reversible isolation. Commit, push, merge, pull request, deployment, external messages, and destructive cleanup need explicit authority. Reuse existing authorization. Ask only for a material unresolved scope, cost, or authority decision.

For worker dispatch, obtain or reuse one approved snapshot through shared/references/model-preview.md. Model and effort defaults come only from shared/model-routing.json. Dispatch, repairs, and resumes retain the exact route, provider controls, tools, privacy, source bindings, budget, fallback triggers, and call, time, concurrency, and recovery limits. A missing required capability blocks that route. Never call models-consensus here.

Use separate persistent contexts for each task, track, and role. Reviewers use a different model and a separate context from every author in their scope. Native handoffs require parent_history: none. Keep native contexts in native_contexts[route_id] and runner sessions in runner_contexts[route_id]. Reconcile pending calls before retry or reconstruction. Context loss preserves counters and routes.

The launcher validates approved input hashes before writes or dispatch. It includes the full contract for first calls and reconstruction, and validates continuation proof before selecting the smaller full or compact input. Derived notes cannot change acceptance. Its complete rendered input limit is 24,000 UTF-8 bytes; a larger approved context_budget needs a reason. Read shared/references/context-packets.md when preparing or recovering a packet.

Select one path

Use the approved policy, not a prose label. Without a bound review_policy, keep task mode. Existing plans and snapshots remain unchanged. A standalone small change uses task mode when its complete contract leaves no product, data, security, public contract, or dependency decision unresolved.

Read only the selected row and its triggered references. Section pointers mean the named section, not the whole manual. Retain already loaded rules without rereading them.

Current workLoad
Implement a feature task classified deferredtask-implementation.md, then deferred-task.md.
Implement a critical feature task or a standalone/legacy tasktask-implementation.md, then task-review.md when implementation evidence is ready.
Review a completed critical or standalone tasktask-review.md.
Complete combined feature review under a bound policyfinal-feature-review.md.
Repair, recheck, or recover a selected pathKeep that path; load Continue the same role in runner-invocations.md, plus shared/references/incremental-review.md for a reviewer recheck or prose addendum.

For a new route selection, load the routing schema, shared/references/task-shaped-model-routing.md, and shared/references/host-model-execution.md. For runner execution, read only Runner policy and receipts. For native execution, read only Native receipts. For job inspection or authorized cleanup, read its Poll and cleanup section. A handoff or successful poll proves execution status only.

Output contract

Return status, acceptance, routes_used, review, changed_paths, and next_action. Include required behavior and captured results, exact approved route IDs and effective receipts, applied or declined findings with evidence, and unresolved risks. Use implemented for provisional feature tasks; otherwise complete, failed, ceiling_hit, or awaiting_human. Critical task review readiness does not accept the feature. Name any pending final gate.

Write the task report under .ai-workflow/impl-review/<session-id>/<task-id>/report.md when available. Link snapshots, original captures, review records and verifier results where applicable, role contexts, reconstruction, changed paths, and remaining risks. Retain role contexts until fixes, rechecks, and evidence obligations finish. Return only a necessary decision, parent integration, or separately authorized delivery as next_action.

Comments

Markdown supported