version-sentinel
Skill6 starsChanged 2 days agoMITView on GitHub
Use when adding, bumping, or changing a dependency in package.json, requirements*.txt, constraints*.txt, pyproject.toml, Cargo.toml, *.csproj, *.fsproj, or *.vbproj. Triggered automatically by version-sentinel's PreToolUse hook (where the host supports hooks) — this skill explains how to satisfy the block and record a version check.
- Installs packages
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- audit-dependenciesSkill
- audit-dependenciesSkill
- version-sentinelSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 18
If a tool call exits 2 with BLOCKED: version-sentinel, you must:
- Distinctive: line 20
Look up the latest version: Use a web search first:
- Distinctive: line 26
If web search is unavailable, fetch the registry URL directly, or consult a docs tool (e.g.
- Distinctive: line 28
Record the check: Run in the terminal:
- Very specific: line 34
Retry with the verified version: If the blocked input omitted the version or used a floating tag such as latest, rewrite it first.
- Distinctive: line 40
If you genuinely intend to install an older version (CVE mitigation or compatibility), record with:
- Very specific: line 44
After a successful online check and current audit, this can pass the hook and is flagged as intentional-pin (not DRIFT) in bash …
- Distinctive: line 48
Don't fake a source URL you didn't actually see.
- Distinctive: line 49
Don't try to bypass the hook with git commit --no-verify or similar — the hook runs on file edits and terminal commands, not on git.
- Very specific: line 50
Don't unset VS_DISABLE without the user's awareness.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Version Sentinel — Workflow
The version-sentinel setup blocks dependency changes until you've verified the package version against its upstream registry. Here's the required flow. (On hosts without hook support, such as Zed, this same flow applies manually — follow it whenever you touch a manifest or run an install command.)
Mandatory startup audit
From the project directory, run the bundled scripts/check-versions.sh at startup and after manifests or lockfiles change. Hooks require successful current project evidence before dependency work. Registry lookup failure blocks; newer-release drift is reported without upgrading dependencies. File inspection, user questions, audit retries, and version recording remain available.
If a required manifest is missing or invalid, ask Daniel before recreating or repairing that exact file. Do not guess dependencies. Only after actual consent, run bash scripts/recover-manifest.sh approve FILE --user-confirmed 'REASON'. This records an attestation; it does not obtain approval or create contents. Online-verify proposed dependencies before writing approved contents, then rerun the audit. Do not use the consent flag without approval.
When you see a BLOCKED message
If a tool call exits 2 with BLOCKED: version-sentinel, you must:
Look up the latest version. Use a web search first:
npm: search"<pkg> latest version site:npmjs.com"pip/pyproject: search"<pkg> latest version site:pypi.org"csproj: search"<pkg> latest version site:nuget.org"cargo: search"<pkg> latest version site:crates.io"
If web search is unavailable, fetch the registry URL directly, or consult a docs tool (e.g. context7) for the package.
Record the check. Run in the terminal:
bash scripts/vs-record.sh <ecosystem> <pkg> <version-you-intend-to-install> <source-url>The source must be the canonical
https://package page or upstream URL, orintentional:followed by a nonempty reason for a deliberate pin. The recorder makes its own successful online upstream request and verifies that the exact intended version exists. Older pins need both online evidence and a reason.- Retry with the verified version. If the blocked input omitted the version or used a floating tag such as
latest, rewrite it first. For example, usenpm install lodash@4.17.21, notnpm install lodashornpm install lodash@latest.
Local paths, URLs, git sources, and workspace dependencies are outside registry-version checks. A NuGet PackageReference without Version may use Central Package Management, so the manifest gate ignores that form.
Intentional non-latest pins
If you genuinely intend to install an older version (CVE mitigation or compatibility), record with:
bash scripts/vs-record.sh <ecosystem> <pkg> <version> "intentional: <brief reason>"After a successful online check and current audit, this can pass the hook and is flagged as intentional-pin (not DRIFT) in bash scripts/check-versions.sh output.
What NOT to do
- Don't fake a source URL you didn't actually see. The recorder validates canonical upstream data; a URL string alone is not evidence.
- Don't try to bypass the hook with
git commit --no-verifyor similar — the hook runs on file edits and terminal commands, not on git. - Don't
unset VS_DISABLEwithout the user's awareness; that's an escape hatch for throwaway sessions, not normal flow.
Audit command
bash scripts/check-versions.sh scans manifests within 4 directory levels of the current directory and reports drift. Run it before tagging a release.
Legacy URL-only records, automatic post-install records, and stale offline responses cannot authorize dependency changes. Private registries are not supported by the current canonical verifier; a reason cannot bypass online verification. Hooks do not intercept arbitrary shell/Python file writes. See docs/enforcement.md for scan scope, recovery trust, and host activation limits.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

