菜单
Loaded by the pipeline-spec-reviewer agent to perform Stage 1 spec-compliance review.
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。
基于 SOC 职业分类
Cut a Superpipelines version release (vX.Y.Z) — verify the 5 version targets agree, draft CHANGELOG + RELEASE-NOTES entries, land them via PR (main is branch-protected), then create the GitHub release + tag and verify. Use when the user asks to cut/ship/publish a release, tag a version, do release work, or write release notes for Superpipelines.
Resolves agent `model_tier:` declarations into concrete platform-specific model strings and effort values via a 5-layer precedence chain, returning a serializable `resolved` object for dispatch. Use when `running-a-pipeline` Phase 0.45 needs to determine the concrete model + effort for each pipeline step on the active platform before dispatch.
Resolves the active runtime tier and provides the canonical DISPATCH contract for all pipeline step execution. Use when an orchestrator skill needs to dispatch pipeline steps — performs tier detection, loads the platform profile, and executes the Tier 2 inline loop when no subagent primitive is available.
Cut a Superpipelines version release (vX.Y.Z) — verify the 5 version targets agree, draft CHANGELOG + RELEASE-NOTES entries, land them via PR (main is branch-protected), then create the GitHub release + tag and verify. Use when the user asks to cut/ship/publish a release, tag a version, do release work, or write release notes for Superpipelines.
Run, resume, or list installed Superpipelines workflows from the registry.
Provides canonical engineering standards for authoring agents, skills, hooks, and pipeline orchestration in Claude Code. Use when authoring or modifying agents, skills, hooks, or pipeline-orchestration artifacts — covers model-tier selection, prompt-cache discipline, frontmatter schemas, and progressive-disclosure rules.
| name | pipeline-spec-reviewer-protocol |
| description | Loaded by the pipeline-spec-reviewer agent to perform Stage 1 spec-compliance review. |
| disable-model-invocation | true |
| user-invocable | false |
When the active platform_profile.capabilities.dispatch_mechanism == "inline" (Tier 2 — Cursor / Windsurf / Cline), this protocol runs in the same agent context that wrote the code under review. Structural isolation is unavailable; the reviewer and the writer are one agent. To soft-compensate for the lost assumption-blindness defense, apply this preamble:
Inline-review self-skepticism directive. You are reviewing your own prior work in this same session. Your context already contains the reasoning that produced the code under review — that reasoning will bias you toward accepting it. Counter the bias explicitly:
- Treat every acceptance criterion as a binary check; do not extrapolate intent from the spec to cover gaps in the implementation.
- When in doubt about whether an AC is met, flag rather than pass. The default is FAIL; PASS requires affirmative evidence.
- List the specific lines / files that satisfy each AC. If you cannot point to lines, the AC is unverified.
- Reject any "the spirit of the AC is met" reasoning. Spec compliance is mechanical, not interpretive.
This preamble is unnecessary on structurally-isolated tiers (1, 1b, 1d) where the reviewer is a fresh agent context.
Evaluate the executor's output against each AC:
Check every modified file for over-build:
files list?<rationalization_table>
| Excuse | Reality |
|---|---|
| "Useful over-build" | Creep violates the contract parallel workers depend on; FAIL always. |
| "Mostly met" | Mostly = Under-build. Functional correctness is not a gradient. |
| "Trust the executor" | Executors cannot review themselves. Verify every AC against the actual output. |
| </rationalization_table> |
${CLAUDE_PLUGIN_ROOT}/skills/sk-write-review-isolation/SKILL.md — Isolation protocol.${CLAUDE_PLUGIN_ROOT}/skills/sk-claude-code-conventions/SKILL.md — Formatting rules.