SuperAPI AI Directory

review-method

Coding & development

View repository

Updated 10-15 09:53

About this skill

Unified review method

Input: variant, target/flags, output root; optional same-snapshot prepared bundle. Output: native report and the project-result envelope. Do not call a second workflow as an executor.

The consuming workflow declares the named reviewer inventory in dispatch.roles and isolated inline workers in its sidecar. Any inline lens, critic, history, hygiene or minimal-fix brief in the references uses the canonical project-protocol:isolated-worker binding supplied by the host harness. A general-purpose label describes that inline task; it is not another agent ID.

  1. Load the senior-review:review-preparation skill, unless a same-snapshot prepared bundle was supplied. Validate its target, run identity and snapshot before reusing it. Complete missing stack selection through review-preparation before dispatch; Validate the prepared knowledge selection's schema, candidate/run/snapshot, content and metadata fingerprints, scope and assigned delivery IDs as well. A missing or stale knowledge binding requires preparation against this exact candidate before dispatch, not reuse of an older-schema or another-run bundle. a reused bundle needs activation evidence or a skip/gap reason for all three stack dimensions. Do not prepare another run or dispatch a reviewer twice.
  2. Read references/<variant>.md in this skill. Dispatch the full prepared selection, including applicable React, TypeScript and platform reviewers, in the same batch. Core role inputs and prompts remain canonical in senior-review:review-quality-gates; stack roles use their canonical provider definitions and preparation's shared inputs. Preserve the prepared knowledge_selection and assign each worker only its scoped knowledge_ids, selection fingerprint and exact catalog load/reference bindings. Each worker loads its assigned knowledge as read-only input, applies it against code evidence, and returns knowledge_usage with scope, worker, hashes, references and application explanation. Account for task-specific knowledge lenses in this same dispatch and delivery ledger. No worker receives peer findings before delivery. Use isolated contexts and the host's supplied coordination mechanism. Account for every selected worker before starting consolidation.
  3. Load the senior-review:review-consolidation skill with all native reviewer results. Do not run the reference's output/fix tail until consolidation finishes.
  4. For code review, read references/code-review-delivery.md only after all core and stack results have passed native consolidation and verification. Team and PR report delivery remains in their consolidation references. Run delivery and any authorized fix steps once. Application subtraction loads senior-review:application-cleanup-method; significant semantic refactors are returned to the caller's development plan and are never delegated to cleanup.

Read-only review preserves the workspace. Before any accepted fix, load project-protocol:project-protocol for execution preflight, authorizations, baseline, candidate gate and owned recovery. Commit and publication are separate authorizations; a request for review is not authorization to publish.