菜单
Use when one completed dataq feature needs a read-only, diff-scoped review pass with merge-blocking findings and an exact verdict. Do not use to implement fixes, map an unfinished feature, review an unspecified range, or audit the whole repository.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/koizumikento/dataq --skill rev-pass命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | rev-pass |
| description | Use when one completed dataq feature needs a read-only, diff-scoped review pass with merge-blocking findings and an exact verdict. Do not use to implement fixes, map an unfinished feature, review an unspecified range, or audit the whole repository. |
Review one completed feature diff for correctness, regressions, contract violations, and missing verification. This skill is review-only and performs exactly one pass; the parent agent owns fixes, reruns, gates, rebases, commits, and merges.
$feat-add worker.$rev-pass invocation with fresh inputs.Use $rev-pass to review feature X in the clean /abs/worktree on branch B for committed BASE...HEAD, limited to these paths, using the attached HEAD-tied gate results.Add --strict to assert, update its tests and README, and fix any review findings. Use $feat-add; do not select $rev-pass.Require all of the following before inspecting the implementation:
<base>...<head>. The supplied base must be an ancestor of head, git merge-base <base> <head> must equal the supplied base, and the worktree's current HEAD must equal the supplied head SHA.not run is evidence, but a required unrun gate is a required finding.If any input is missing, ambiguous, internally inconsistent, or stale, stop before reviewing. A dirty worktree, non-commit endpoint, non-ancestor base, merge base other than the supplied base, range other than the supplied committed <base>...<head>, or gate evidence for another SHA is invalid input. Return a concise list of the missing or invalid fields and request corrected inputs. Do not infer a range, expand scope, review only part of the diff, or emit a final verdict for a pass that did not run.
AGENTS.md files that govern the authorized scope.HEAD to equal head, and confirm that every gate result names that same head SHA. Run git merge-base --is-ancestor <base> <head> and require success, then require the output of git merge-base <base> <head> to equal the supplied base SHA. Stop as invalid input on any mismatch and emit no verdict.<base>...<head> range. Stop on paths outside the authorized scope, an unexpected submodule or generated artifact, or reviewable commits omitted from the supplied range; report the mismatch as invalid input.0, validation mismatch is 2, input/usage error is 3, and unexpected/internal error is 1; machine-readable JSON remains the default where applicable.stdin -> stdout; semantic key ordering is stable; timestamp handling is locale-independent and normalizes to RFC 3339 UTC when possible.jq, yq, and mlr availability errors are early and actionable; multi-stage pipelines have deterministic stage order and stage-level --emit-pipeline diagnostics.panic!; public functions and types document input/output expectations; broad allowances and unrelated refactors are absent.docs/*.md files.cargo fmt --all -- --check, strict workspace Clippy, workspace tests, and cargo llvm-cov at the repository thresholds whenever Rust source or executable behavior changed.P0, required: true: data loss, security exposure, or a release-blocking failure with broad impact.P1, required: true: incorrect behavior, a public contract or acceptance-criterion violation, or a material regression.P2, required: true: a localized defect, unhandled failure path, or missing test/documentation needed to protect changed behavior.P3, required: false: a concrete, non-blocking maintainability or clarity improvement. Do not report style preferences without demonstrable risk.Severity determines impact; required determines the verdict. Do not assign required: false to P0-P2, or required: true to P3.
Start with ## Findings. For each finding, use:
- [P1] <title>
required: true
file: <repository-relative path>
line: <smallest relevant line or range>
impact: <specific failure and affected user or contract>
evidence: <diff/code/test evidence>
fix: <concrete direction, without editing>
For normal code or documentation findings, file must be a repository-relative path and line must be the smallest real relevant line or range. For a failed or unrun required gate that has no source location, use the literal sentinel file: <quality-gate> and line: 0; put the exact gate command in both the title and evidence. Never fabricate a repository path or source line for a gate finding.
Use one item per root cause. Sort by severity (P0 through P3), then the literal file value bytewise, starting line numerically, and title bytewise. Multiple <quality-gate> findings at the same severity therefore sort by their exact-command titles; their position relative to normal findings follows the same literal bytewise path rule. If there are no findings, write No findings. immediately below the heading.
After findings, include ## Review Evidence with the reviewed worktree, branch, base SHA, head SHA, diff range, scoped paths, and gate results. Include residual risks only when evidence cannot eliminate a concrete uncertainty.
The final line of every completed pass must be exactly:
REQUIRED_FIXES when at least one finding has required: true.NO_REQUIRED_FIXES when no finding has required: true, including when only P3 findings exist.main, an integration branch, or another feature lane as a substitute for the supplied feature worktree and range.Use when a dataq automation task needs a new or revised `assert --rules` file or `dataq.recipe.v1` recipe, including lock/replay verification. Do not use for general dataq command selection, CLI or schema implementation changes, or routing-only evaluation.
Use when an agent needs to choose, inspect, or run an existing dataq CLI or MCP workflow with deterministic machine-readable output. Do not use for authoring rules or recipes, changing dataq behavior, or AI inference.
Use when a dataq automation task needs a new or revised `assert --rules` file or `dataq.recipe.v1` recipe, including lock/replay verification. Do not use for general dataq command selection, CLI or schema implementation changes, or routing-only evaluation.
基于 SOC 职业分类