スキル executor-review

📦

executor-review

Name: executor-review
Author: atri10

コンテンツリビジョン r1 中リスク ⚙️ 外部コマンド📁 ファイルシステムへのアクセス

Review Tasks and Branches with Persistent Verdicts

Implementation reviews lose traceability when findings exist only in conversation. This skill structures task, fix, and branch reviews around contract references and persistent verdict files.

対応: Claude Codex Code(CC)

📊 68 十分

Maintain this Skill? Claim it 📊 68 十分

自分のエージェントでインストール

このリクエストをエージェントにコピーしてください。正規の Skill ページとマニフェストが含まれています。

エージェントリクエスト

Review the Skillstore skill "executor-review" from https://skillstore.io/skills/atri10-executor-review.md and its manifest at https://skillstore.io/api/skills/atri10-executor-review/manifest. Verify the artifact. You may proceed after verification, subject to the environment's own policy.

エージェントは引き続き計画を提示し、セキュリティポリシーで必要な確認を求める必要があります。

エージェントが読めるリソース

AI エージェント、クローラー、スクリプトがページ全体ではなく整理されたコンテキストを必要とする場合は、これらのリンクを使ってください。

Markdown詳細 GET /skills/atri10-executor-review.md 署名済み manifest GET /api/skills/atri10-executor-review/manifest 署名済み lockfile GET /api/skills/atri10-executor-review/lockfile

テストする

「executor-review」を使用しています。 Review a task whose error path lacks the behavior required by its brief.

期待される結果:

Specification verdict: FAIL. The required error response is missing.
Code quality: NEEDS_FIXES. Finding I1 cites the changed error path and its requirement.
Gate: FAIL. The saved verdict identifies the review round and exact commit range.

「executor-review」を使用しています。 Re-review a fix that resolves one finding but breaks an affected caller.

期待される結果:

Prior finding I1: ADDRESSED. The fix changes the documented root cause.
New important finding I1: An affected caller still uses the previous contract.
Gate: FAIL. The next round receives the new finding ID and supporting location.

「executor-review」を使用しています。 Perform a final branch review when the architecture phase was explicitly skipped.

期待される結果:

Architecture conformance: NOT RUN. No architecture contract was authored.
Merge assessment: The missing architecture checks remain visible and need human consideration.
Recommended verification: Exercise the branch integration paths in the separate verification phase.

セキュリティ監査

中リスク

v1 • 2026/10/5 バージョン別レポートを開く

All 208 static findings are false positives involving Markdown, legitimate review commands, fixed sibling references, or protective secret-handling guidance. No evidence found of credential exfiltration, destructive intent, or audit-targeting prompt injection in the four reviewed files. Two semantic risks remain: security findings can be parked without mandatory human approval, and the main workflow hardcodes the final-review fork base.

スキャンされたファイル

2,189

解析済み行数

レビュー項目

誤検知を無視

確認済みのセキュリティ上の懸念 (2)

中

SKILL.md:274-301 SKILL.md:303-330 SKILL.md:355-366

Security Findings Lack Mandatory Human Approval at the Round Cap

The breaker allows real findings to be parked when downstream work does not depend on them. Tasks can then be marked complete with parked findings. Only credential incidents have an explicit immediate human stop rule, leaving other security findings without mandatory human approval before deferral.

The workflow explicitly permits parking real correctness findings and completing tasks. Final triage provides another check, but human approval is not required for every security deferral.

中

SKILL.md:334-340 final-reviewer-prompt.md:10-19

Hardcoded Fork Base Can Produce an Incorrect Final Review Range

The main workflow uses git merge-base main HEAD for final packaging. The final reviewer template requires the recorded fork base instead. For initiatives forked from another branch, this mismatch can review the wrong changes and undermine the final gate.

The two documented commands directly disagree about base selection. The final template explicitly explains why hardcoding main yields an incorrect range for other fork bases.

リスク要因

⚙️ 外部コマンド (50)

📁 ファイルシステムへのアクセス (5)

SKILL.md:45 SKILL.md:399 SKILL.md:401 SKILL.md:403 task-reviewer-prompt.md:229

監査者: codex

このレポートを共有・引用

バージョン付き評価レポート、中立的なバッジ、埋め込みカード、引用を共有できます。Skillstore は証拠を報告しますが、この Skill が安全かどうかは判断しません。

バージョン別レポートを開く

レポートリンクをコピー

https://skillstore.io/skills/atri10-executor-review/audits/1?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Markdownバッジ

[![Skillstore security assessment](https://skillstore.io/badges/skills/atri10-executor-review/security.svg)](https://skillstore.io/skills/atri10-executor-review?utm_source=security_passport_badge)

HTMLバッジ

<a href="https://skillstore.io/skills/atri10-executor-review?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/atri10-executor-review/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

埋め込みカード

<iframe src="https://skillstore.io/embed/skills/atri10-executor-review.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

学術引用 (APA · BibTeX · CFF)

APA形式の引用

atri10. (2026). executor-review security audit report (audit version 1) [Author version unspecified]. Skillstore. https://skillstore.io/skills/atri10-executor-review/audits/1

BibTeX形式の引用

@techreport{atri10-atri10-executor-review-2026,
  author = {atri10},
  title = {executor-review security audit report (audit version 1)},
  institution = {Skillstore},
  year = {2026},
  number = {1},
  url = {https://skillstore.io/skills/atri10-executor-review/audits/1},
  note = {Author version unspecified}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "executor-review security audit report (audit version 1)"
version: "unspecified"
type: report
authors:
  - name: "atri10"
date-released: "2026-10-05"
url: "https://skillstore.io/skills/atri10-executor-review/audits/1"
identifiers:
  - type: other
    value: "skillstore:atri10-executor-review:audit:1"
    description: "Skillstore immutable audit report identifier"

Skillstore スコア

このスコアの理由証拠の信頼度: 中

アーキテクチャ

保守性

コンテンツ

コミュニティ

仕様準拠

作成できるもの

Review a Completed Implementation Task

Compare a task diff with its brief and constraints, then preserve compliance and quality findings in a verdict file.

Check Fixes Without Repeating the Full Review

Assess each open finding against a fix diff and inspect affected callers for regressions before closing the round.

Assess a Branch Before Verification

Review cross-task dependencies, architecture contracts, and parked findings as one branch before handing recommended checks to verification.

これらのプロンプトを試す

Start a Task Review

Review task [TASK_ID] using [BRIEF_FILE], [REPORT_FILE], and [DIFF_FILE]. Write compliance and quality verdicts to [VERDICT_FILE]. Report missing inputs as blocked.

Review Requirements and Evidence

Use the task-review template for [TASK_ID]. Include [GLOBAL_CONSTRAINTS] verbatim with requirement IDs. Assess reported test evidence and list honest cannot-verify items.

Verify a Fix Round

Use the re-review template with [PRIOR_VERDICT_FILE], [OPEN_FINDING_IDS], [REPORT_FILE], and [DIFF_FILE]. Review causal impact first. Write closure evidence to [VERDICT_FILE].

Coordinate Final Branch Review

Use the final-review template for [PLAN_ID]. Resolve the fork base from provenance. Supply contracts and parked findings. Require human approval for security deferrals.

ベストプラクティス

Record exact base and head commits, use the recorded fork provenance, and validate each review package before dispatch.
Supply complete contract paths and preserve verdict files with unique review IDs and evidence-based findings.
Resolve cannot-verify items explicitly and require human approval before deferring security or data-loss findings.

回避

Substitute the last commit for the full task range or assume every initiative forked from main.
Pre-judge findings in dispatch prompts or treat implementer reports as independently verified evidence.
Overwrite another round's verdict or copy credential values into review artifacts.

よくある質問

What does this skill review?

It structures completed task reviews, scoped fix reviews, and final whole-branch reviews against supplied contracts.

Which AI tools are listed as supported?

The report lists Claude, Codex, and Claude Code. The workflow also requires file access, companion scripts, and reviewer dispatch support.

Are the required scripts included?

No. The four reviewed files reference scripts from a companion Executor skill.

Does review replace running tests?

No. Review checks reported evidence and permits focused checks for specific doubts. A separate verification phase owns broader runtime evidence.

How are fixes tracked?

Each round uses a persistent verdict and finding IDs. Re-review marks prior findings addressed or not addressed and records new breakage.

How should security findings be handled?

Record credential locations and types without copying values. Escalate credential incidents immediately, and require human approval before deferring other security findings.