Docs & Knowledge · daymade/claude-code-skills
marketplace-health-check
Run a full 6-dimension health check of this Claude Code skills marketplace repo — code/script safety, documentation/SSOT consistency, security/PII leaks, open-PR triage, open-issue triage, and marketplace-manifest integrity — via a parallel fan-out Dynamic Workflow, then verify the serious findings and report them by priority. Use this whenever the user asks to check the repo, run a health check, do a full sweep/audit before a release, 全面体检, 检查仓库状态, 看看仓库健康吗, 审计一下仓库, or asks whether the PRs / issues / docs / versions / PII are in good shape across the board — even if they never say the word "workflow". Reach for it for any broad "is this whole repo OK" request, not just one-file checks.
L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandInstall DepsNetwork AccessWrite FileDual-pass Audit
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.
1How It Works & Mechanism
Execution uses a 6-way parallel fan-out: six independent inspectors each handle one dimension (code/script safety, doc SSOT consistency, security/PII, open PR triage, open issue triage, marketplace manifest integrity), running in parallel at the same time.
marketplace-health-check/SKILL.md
Six independent inspectors cover, in parallel:
There are no dependencies across the six dimensions, allowing complete parallelism—this is the sole source of this skill's performance gain compared to a single agent serially scanning the entire repo; each inspector carries only its own prompt segment, preventing context pollution.
Parallelism is implemented directly in code rather than just suggested: after phase('Inspect') in workflow.js, parallel() spawns an agent() for each of the six dims, with each agent bound to the same CHECK_SCHEMA structured output contract.
marketplace-health-check/scripts/repo-health-check.workflow.js
const checks = await parallel(dims.map(d => () => agent(COMMON + d.prompt, { label: 'check:' + d.key, phase: 'Inspect', schema: CHECK_SCHEMA })))
How it works: each item in the dims array represents a dimension (key + a long prompt), parallel orchestrates concurrent dispatch, and agent(..., {schema}) enforces that each inspector yields an identically shaped JSON, giving the main agent six aggregatable results.
Must run inline; context: fork is prohibited—forked subagents cannot spawn further subagents or launch workflows, so running forked will cause the entire fan-out to fail silently.
marketplace-health-check/SKILL.md
**This skill must run inline (no `context: fork`).**
SKILL.md explicitly documents this in the section "Why a workflow — and why it MUST run inline"; it is a hard constraint coupled to the execution runtime, not a stylistic preference.
The script is passed entirely inline via the script parameter of the Workflow tool, rather than having the host look up the file path.
marketplace-health-check/SKILL.md
Read the bundled script and launch it **inline via the `script` parameter**
The matching invocation pattern is provided in the SKILL.md code block: Workflow({ script: <full contents of scripts/repo-health-check.workflow.js>, args: { repo: "<owner/name>", scale: "<one-line summary from Step 1>" } })—eliminating the variable of where the skill is installed.
Runs a quick initial scout to collect repo scale (repo name, skill count, open PR/issue counts, manifest version, HEAD, latest release), sharing this as args across the six inspectors; if omitted, each inspector rediscovers it independently.
marketplace-health-check/SKILL.md
gh repo view --json nameWithOwner,stargazerCount,isPrivate | jq -c .
Corresponds to the ctx logic in workflow.js: const repoName = ctx.repo || … and const scale = ctx.scale || 'NOT pre-supplied …', as well as the comment args is OPTIONAL — if omitted, each inspector self-discovers the repo scale first.; SKILL.md also requires confirming isPrivate: false before treating PII as a release risk.
The output contract is strictly enforced by schema: each dimension returns dimension/health/summary/findings[]/stats, where each finding must include severity (critical|high|medium|low|info), title, detail, location, and recommendation.
marketplace-health-check/scripts/repo-health-check.workflow.js
severity: { type: 'string', enum: ['critical', 'high', 'medium', 'low', 'info'] },
In the same schema, health is restricted to four levels (good|minor-issues|needs-attention|critical) with additionalProperties: false—allowing the six results to be directly assembled into a health table while preventing inspectors from hallucinating arbitrary fields.
Strict output discipline is the core of this skill: inspector conclusions are explicitly treated as hypotheses that cannot be passed through unverified; the executor must manually re-verify them before reporting.
marketplace-health-check/SKILL.md
**Agent findings are HYPOTHESES, not conclusions. Never relay them verbatim.**
SKILL.md requires verifying each high/critical finding using a single quick command (grep for leaked values, sed -n to inspect broken lines, gh repo view to check claimed states), followed by filtering through four criteria: probability / cost / real scenario / verifiable.
Reports are layered by priority and evidence confidence rather than dumping 25 raw findings onto the user.
marketplace-health-check/SKILL.md
- **🔴 Must-fix** — each VERIFIED high/critical, with exact location + a concrete fix
Structure: one-line conclusion + 6-dimension health table → 🔴 Must-fix → 🟠 Backlog → 🟢 Optional → 💡 Key insights; each tagged with ✅ real / ⚠️ partly / ❌ false-alarm, as SKILL.md explicitly notes that most raw agent outputs are noise.
Operates strictly in read-only mode throughout (inspectors must not modify files, comment on PRs/issues, or push), and states the estimated cost up front before starting.
marketplace-health-check/scripts/repo-health-check.workflow.js
This is READ-ONLY: do NOT modify files, comment on PRs/issues, or push. gh CLI is authenticated.
Cost is quantified in SKILL.md: ~15-20 min, ~400-500k output tokens with the requirement to tell the user the cost up front; subsequent remediation is left to the repository owner, with the skill making no automated changes or comments.
2Core Capabilities
01Parallel orchestration of 6-dimension repo health checks (Workflow script + 6×agent parallel fan-out)
02Structured finding outputs (4-tier health / 5-tier severity / mandatory location and recommendation)
03Individual open PR triage: worth-merging / needs-changes / low-quality-or-spam / SHOULD-DECLINE, plus checking for mandatory version bumps and CHANGELOG updates often missed by external PRs
04Open issue categorization (real-bug / skill-request / cross-list-or-promotion / question) with proactive investigation into broken-install-command bugs
05Marketplace manifest integrity validation: executing check_marketplace.sh and check_doc_skill_lists.py to reconcile orphan SKILL.md files, suite member registrations, and version/count alignments
06Semantic security/PII auditing: keyword-agnostic leak detection (real personal names, private domains), missing .security-scan-passed marker tracking, and human-readable case file verification
07Counter-Review rule set: anti-target evaluation / git history exposure / scan marker necessity vs. sufficiency / mandatory versioning / promotion default rejection
08Final reports layered by priority + evidence confidence tier (🔴/🟠/🟢 + ✅/⚠️/❌ + insights)
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| cli | gh（GitHub CLI） |
| cli | jq |
| cli | git |
| cli | bash（执行仓库自带检查脚本） |
| cli | python3 |
| cli | uv（跑 skill-creator 的 security_scan 模块） |
| cli | claude plugin validate（经 check_marketplace.sh 调用） |
| cli | gitleaks（经 security_scan.py 调用） |
| cli | grep / find / sed（检索与逐行核对） |
| api | GitHub REST API（经 gh CLI：PR / issue / release / repo 元数据，只读） |
4Recommended Governance & Evidence
Suggested Controls for Safe Operation
For Agent Runtime
Graded/per-use authorization, sandbox recommended
For Administrators
Approval workflow, rollback support, full audit
Evidence & Triggers (File Paths & Excerpts)
- scripts/: Executes local scripts or commands: marketplace-health-check/scripts/repo-health-check.workflow.js — via Workflow
- scripts/ 或 SKILL.md: Declares dynamic installation of external language package dependencies
- 网络端点: Declares external network access: access to GitHub REST via gh CLI (api.github.com, default gh endpoint, domain not explicitly in repo text): gh repo view /
- 本地工作目录: File I/O and artifact output: read-only by design (COMMON prompt and script comments state READ-ONLY: inspectors must not modify files, c
Audit Notes: This is not a pure prompt skill: the Workflow script spawns 6 inspectors equipped with shell and gh execution permissions, executing read-only queries against api.github.com using logged-in GitHub credentials, and explicitly invoking repository scripts (check_marketplace.sh / check_doc_skill_lists.py / security_scan.py). There is no third-party data exfiltration, secret file harvesting, or destructive file modification, so it does not reach the orange tier. However, because it runs scripts and makes credentialed network calls, it sits higher than pure read-only green/blue tiers. Residual risks include: 1) code-level side effects where security_scan.py modifies markers (read-only promise relies on git checkout cleanup); 2) read-only behavior is constrained by prompt instructions rather than hard OS/tooling permissions, meaning inspectors technically hold gh write tokens; 3) prompt injection surface (untrusted repository files and PR/issue bodies fed into privileged agent contexts) represents the primary uncertainty.
5Second-Pass Verification
- [ok] 六维清单与代码 dims 一一对应 — SKILL.md 列的六维（代码/脚本安全、文档 SSOT、安全/PII、开放 PR、开放 issue、marketplace 清单）与 workflow.js 的六个 key（code-and-script-safety / doc-consistency / security-pii / pull-requests / issues / marketplace-integrity）逐项对应，dims 数组经计数恰为 6 项，无多无少。
- [ok] parallel fan-out 是真实实现而非修辞 — 源码含 phase('Inspect') 与 parallel(dims.map(d => () => agent(COMMON + d.prompt, { label: 'check:' + d.key, phase: 'Inspect', schema: CHECK_SCHEMA })))，description 里「parallel fan-out Dynamic Workflow」无夸大。
- [ok] 必须 inline 的运行约束 — SKILL.md 逐字存在该段，且理由（forked subagent 不能 spawn subagents 或启动 workflow）与 workflow 脚本调用 agent()/parallel() 的宿主能力要求自洽。
- [ok] 外部 CLI 调用点真实存在 — gh、jq、git、bash、python3、uv、claude plugin validate、gitleaks 均在 SKILL.md 或 workflow.js 逐字可定位；gitleaks 仅出现在 methodology 对 security_scan.py 实现的描述里，属间接调用（该脚本确实存在于 daymade-skill/skill-creator/scripts/security_scan.py）。
- [unlocatable] gh 网络端点域名 — 仓库内检索 api.github.com / github.com 零命中：skill 只写 gh CLI is authenticated. 与各 gh 子命令，端点域名按 gh CLI 的默认行为推断为 api.github.com，无法在源码内取证，已在 external_deps 与 security.network_calls 中标注该推断。
- [discrepancy] 只读承诺 vs 实际副作用 — 脚本注释、COMMON prompt 与 SKILL.md 三处 READ-ONLY 声明针对「文件修改/评论/push」，但 methodology 自述 security_scan.py MUTATES the marker file，并要求 git checkout 还原。故第一遍若把本 skill 概括为「零写入」属过强；准确说法是「对代码与协作面无写入，但被点名执行的扫描器有一条 marker 写副作用，靠人工还原」。
- [ok] 资产引用路径真实存在 — workflow.js 引用的 check_marketplace.sh、check_doc_skill_lists.py 与 security_scan.py 在 pin commit 下均真实存在；SKILL.md 提到「repo 根、bundles 之外的 decline policy 参考文档」对应仓库根 references/promotion-policy.md，CLAUDE.md 亦以 references/promotion-policy.md 指向它。
- [unlocatable] 成本声明可验证性 — 「~15-20 min, ~400-500k output tokens」来自该 skill 蒸馏时的实测会话，仓内无基准、日志或脚本可复算，只能视为作者经验值。
6Verdict & Scenarios
Best for: Ideal for pre-release comprehensive health checks on mature multi-suite repositories using marketplace.json as SSOT and maintaining open PR/issue queues (e.g., daymade/claude-code-skills); also serves as an exemplary template for other audit skills adopting a 'multi-branch fan-out + structured schema + Counter-Review' workflow (requires updating repo paths and scripts in dims).
Not for: Not suitable for inspecting single files or individual skills (as defined in description: not just one-file checks); not directly portable to other repos without modification (dimension prompts are tied to this repo's tooling); not suitable for environments lacking gh authentication or a GitHub remote (PR/issue/release dimensions will fail); not intended for users expecting automated fixes or bot replies to contributors, as docs explicitly delegate remediation and communication to the repository owner.
Install Agent-Ready
① Verified Mirror 2026-09-15
Option A · Direct DownloadDownload marketplace-health-check.tar.gz
sha256:
Option B · Copy Guide for Agent066980771e4106d8…Agents automatically download and verify mirrors from this JSON guide.
② Upstream GitHub
GitHub Repo ↗
Pinned to commit d5c4678cb5.