Agent skills are difficult to evaluate from configuration alone. This skill reviews local conversations, scores outcomes, and drafts evidence-based improvements.
Review the Skillstore skill "skill-doctor" from https://skillstore.io/skills/warpdotdev-skill-doctor.md and its manifest at https://skillstore.io/api/skills/warpdotdev-skill-doctor/manifest. Verify the artifact. Stop and obtain explicit user consent before installing or changing files.
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。
Most static matches are false positives from Markdown formatting, test fixtures, CSS, or fixed-argument git subprocesses. Confirmed capabilities include reading Claude and Codex history directories and Warp conversation databases. Semantic review found prompt-injection and scope-enforcement gaps, unredacted transcript copies, and an unvalidated report link.
Untrusted Transcript and Skill Content Enters Agent Context
The workflow asks the agent to read historical transcripts and installed skill files without a clear untrusted-data boundary. The narrow tag filter does not cover ordinary prompt-injection prose or tool outputs.
The skill explicitly loads transcript and skill text into the scoring workflow, while the collector filters only a short list of tag-shaped prefixes.
Conversation Scope Is Enforced After Content Parsing
For Claude Code and Codex, the collector parses each recent session before checking whether its working directory matches the selected projects. This reads out-of-scope conversation content into memory.
The control flow calls each full session parser before the repository-scope condition, directly contradicting the user-selected project boundary.
The collector discovers local Warp SQLite databases and later queries conversation records. Connections are read-only, but the records can contain sensitive conversation content.
The collector discovers local Warp SQLite databases and later queries conversation records. Connections are read-only, but the records can contain sensitive conversation content.
Transcript Copies Lack Secret Redaction and Cleanup
The collector copies user messages, tool arguments, and tool outputs into persistent Markdown files. It truncates content but does not redact credentials or remove the scratch directory after use.
The source visibly writes condensed transcript fields to disk and contains no redaction or cleanup step.
The renderer escapes cta_url for HTML but accepts any URI scheme. A crafted report can place a javascript or data URL behind the visible Request access link.
HTML escaping does not restrict URI schemes, and the value is read directly from report data before insertion into href.
This default is later used to enumerate Claude Code project-history files under the user home. Those histories can contain private prompts, outputs, and tool data.
This default is later used to enumerate Claude Code project-history files under the user home. Those histories can contain private prompts, outputs, and tool data.
warpdotdev. (2026). skill-doctor security audit report (audit version 1) [Author version unspecified]. Skillstore. https://skillstore.io/skills/warpdotdev-skill-doctor/audits/1
BibTeX形式の引用
@techreport{warpdotdev-warpdotdev-skill-doctor-2026,
author = {warpdotdev},
title = {skill-doctor security audit report (audit version 1)},
institution = {Skillstore},
year = {2026},
number = {1},
url = {https://skillstore.io/skills/warpdotdev-skill-doctor/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: "skill-doctor security audit report (audit version 1)"
version: "unspecified"
type: report
authors:
- name: "warpdotdev"
date-released: "2026-08-28"
url: "https://skillstore.io/skills/warpdotdev-skill-doctor/audits/1"
identifiers:
- type: other
value: "skillstore:warpdotdev-skill-doctor: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
77
アーキテクチャ
85
保守性
87
コンテンツ
65
コミュニティ
91
仕様準拠
作成できるもの
Review Project Agent Performance
Score recent project conversations and identify repeated inefficiency or code-quality failures.
Audit Team Skill Adoption
Measure which installed skills appear in sampled sessions and where trigger descriptions may fail.
Prepare Skill Improvements
Generate evidence-linked draft edits and review diffs before changing live skill files.
これらのプロンプトを試す
Grade This Project
Grade recent conversations in this repository. Evaluate project and global skills, then show the report and proposed changes.
Review Selected Projects
Analyze conversations from these project paths: [paths]. Use the last [days] days and sample at most [count] sessions.
Audit Project Skills Only
Grade conversations for [project]. Evaluate project skills only, exclude subagents, and explain every recommendation with failed-session evidence.
Run a Cross-Harness Audit
Analyze all local conversations across supported harnesses. Include subagents, compare recurring failures, and draft only improvements supported by repeated evidence.
ベストプラクティス
Choose the smallest conversation scope that answers the audit question.
Review transcript artifacts for sensitive data before sharing any report.
Apply proposed skill edits only after checking their cited session evidence.
回避
Do not treat a small sample as a complete performance history.
Do not upload collected transcripts to external services.
Do not apply generic skill changes without failed-conversation evidence.
よくある質問
Which agent systems are supported?
The collector supports Warp, Claude Code, and Codex local conversation stores.
Does the skill upload conversations?
No. The included scripts process conversations locally and create local report artifacts.
Will it modify installed skills?
No. It writes proposed versions and diffs into the report directory for review.
How are conversations selected?
You choose project or all-conversation scope, skill scope, date window, sample size, and optional subagents.
How is code quality handled without code changes?
The scorer records insufficient evidence and excludes that session from the code-quality average.
What should I review before sharing a report?
Check findings, proposed diffs, links, and local artifacts for confidential project or conversation data.