一键导入
deep-review-workflow
Internal cross-runtime review pipeline used by the public deep-review skill.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Internal cross-runtime review pipeline used by the public deep-review skill.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
Alternate independent review and evidence-based response until convergence on Claude Code or Codex.
Public cross-runtime entrypoint for independent review, initialization, and evidence-based review response.
코드 리뷰 피드백 수신 시 증거 기반 대응 프로토콜. 맹목적 동의를 차단하고, 기술적 검증 후 증거와 함께 수락/반박/구현한다. deep-review verdict 또는 외부 리뷰(PR 코멘트, 동료 리뷰) 모두에 적용.
基于 SOC 职业分类
| name | deep-review-workflow |
| description | Internal cross-runtime review pipeline used by the public deep-review skill. |
| user-invocable | false |
This internal skill is loaded only by the public review branch. Public users
invoke $deep-review:deep-review or its /deep-review Claude adapter.
references/runtime-dispatch.md — capability-based entry and reviewer matrixreferences/review-execution.md — executable review pipelinereferences/review-criteria.md — six review lenses and severity doctrinereferences/codex-integration.md — Codex roles and synthesisreferences/agy-integration.md — privacy-gated agy rolereferences/ultracode-integration.md — optional six-lens Claude fan-outreferences/recurring-findings-export.md — Stage 5.5 exportreferences/init-setup.md — public init terminal branchreferences/contract-schema.md and references/report-format.md — data contractsThe reviewed runtime-reference file map is the list above. Every executable
review or loop path uses .mjs/.js helpers or direct host tools.
Resolve plugin_root in this order: generic PLUGIN_ROOT, compatibility
CLAUDE_PLUGIN_ROOT, then the current runtime module location. The generic
override never identifies the host. Invoke
node {plugin_root}/hooks/scripts/detect-environment.mjs --cwd {project_root} --format json
and use its absolute plugin_root value for every later plugin path.
detect-environment.mjs,
build-change-files.mjs, direct Git host commands, and direct file reads.build-reviewer-payload.mjs.runtime-dispatch.md, dispatch all
eligible roles, and enforce read-only fingerprints.report-format.md.mutation-protocol.mjs.wrap-recurring-findings-envelope.js.The detailed stages and failure behavior live only in review-execution.md.
review_model is a non-empty installed Claude model alias; fable is a valid
example. Treat the value as opaque and forward it unchanged to the named
Claude agent or run-claude-reviewer.mjs. Unknown or unavailable aliases fail
that role visibly and never fall back to a different model identity.
The remaining top-level keys retain their existing meaning:
review_model: fable
codex_notified: false
agy_notified: false
agy_enabled: true
agy_sensitive_acked_fingerprint: ""
agy_sensitive_acked_at: ""
agy_fingerprint_mode: hybrid
last_review: null
app_qa:
last_command: null
last_url: null
Patch only the intended top-level key through the Node config writer or a precise host edit. Preserve unknown user keys.