agentleFS
KSEGITKSEGIT/Version-Sentinel/AGENTS.md

Version-Sentinel

AGENTS.md6 starsChanged 2 days agoMITView on GitHub

Version Sentinel is a guardrail that prevents stale, hallucinated, or compromised dependency versions from reaching your manifests. Before any dependency is added, bumped, downgraded, or installed, the agent must verify the intended version against its upstream registry and record that check. Checks are stored in .version-sentinel/checks.json and expire after a freshness window (default 24h). Before dependency work, complete the online startup audit. Missing or invalid manifests require Daniel's explicit approval before recovery. The recorder now performs a successful online upstream…

  • Installs packages
  • Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.

    Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.

  • In this repository6 other files it gives its agents.

The file

We split this AGENTS.md into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If a hook blocks an edit or install (exit 2, BLOCKED: version-sentinel), or if you are about to change any dependency on a platform without …

    line 12
  2. Very specific:

    Retry with the verified version: If the blocked edit or command omitted the version or used a floating tag such as latest, replace it with …

    line 32
  3. Distinctive:

    Never fabricate a source URL you did not actually consult, and never bypass the record step to force a dependency change through.

    line 34

≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
AGENTS.md

Version Sentinel — Cross-Agent Instructions

Version Sentinel is a guardrail that prevents stale, hallucinated, or compromised dependency versions from reaching your manifests. Before any dependency is added, bumped, downgraded, or installed, the agent must verify the intended version against its upstream registry and record that check. Checks are stored in .version-sentinel/checks.json and expire after a freshness window (default 24h).

Mandatory startup audit and recovery

Before dependency work, complete the online startup audit. Missing or invalid manifests require Daniel's explicit approval before recovery. The recorder now performs a successful online upstream check itself; URL-only, offline, and post-install records do not authorize changes. Older pins still need online verification and a reason. Follow the enforcement workflow for audit, recovery, and interception limits.

Required workflow when a dependency change is blocked

If a hook blocks an edit or install (exit 2, BLOCKED: version-sentinel), or if you are about to change any dependency on a platform without hooks:

  1. Look up the latest version on the upstream registry via web search or fetch:
    • npm: https://www.npmjs.com/package/<pkg>
    • pip / pyproject: https://pypi.org/project/<pkg>/
    • cargo: https://crates.io/crates/<pkg>
    • csproj (NuGet): https://www.nuget.org/packages/<pkg>
  2. Record the check:

    bash scripts/vs-record.sh <ecosystem> <pkg> <version> <source-url-or-intentional:reason>

    The source must be the canonical https:// registry URL you actually consulted, or intentional:<reason> for deliberate pins that also pass online verification (e.g. CVE lock, compatibility). Example:

    bash scripts/vs-record.sh npm lodash 4.17.21 https://www.npmjs.com/package/lodash
    bash scripts/vs-record.sh csproj Serilog 3.1.1 "intentional: CVE lock pending audit"
  3. Retry with the verified version. If the blocked edit or command omitted the version or used a floating tag such as latest, replace it with the exact version you recorded. For example, retry npm install lodash@4.17.21, not npm install lodash or npm install lodash@latest.

Never fabricate a source URL you did not actually consult, and never bypass the record step to force a dependency change through.

Audit

bash scripts/check-versions.sh

Scans manifests within 4 directory levels of the current directory, compares each dependency against its upstream registry, and reports drift. Intentional pins show as intentional-pin, not DRIFT. Run before tagging a release.

Supported manifests

package.json (npm/pnpm/yarn/bun), requirements*.txt, constraints*.txt, pyproject.toml (pip, Poetry, uv), Cargo.toml, and *.csproj / *.fsproj / *.vbproj (NuGet).

Local paths, URLs, git sources, and workspace dependencies do not name a registry version, so the gate ignores them. A NuGet PackageReference without a Version may use Central Package Management from an ancestor Directory.Packages.props; the manifest parser also ignores that form. It still blocks floating NuGet versions such as Version="*".

Escape hatch

VS_DISABLE=1 makes all hooks no-op (block, prereq warning, auto-record). Use only for throwaway sessions; do not set it without the user's awareness.

Platform behavior

  • On platforms with hooks (Claude Code and derivatives such as z.ai's ZCode / GLM Coding Plan, Kimi Code, OpenAI Codex), the workflow above is enforced automatically via PreToolUse hooks.
  • On platforms without hooks (e.g. Zed), you MUST follow the same workflow voluntarily before any dependency change: look up the version, run bash scripts/vs-record.sh ..., then make the edit.

Project structure

  • plugin.json, .claude-plugin/ — Claude Code plugin manifest and metadata.
  • kimi.plugin.json — Kimi Code plugin manifest (root); kimi-marketplace.json — Kimi marketplace catalog.
  • gemini-extension.json, GEMINI.md — Gemini CLI extension manifest and context; hooks/gemini-hooks.json + platforms/gemini/setup.sh — Gemini hook wiring.
  • .codex-plugin/plugin.json — OpenAI Codex plugin manifest (points at hooks/codex-hooks.json: the same wiring without the if gating, which Codex ignores — sharing the gated file made it run every handler in each group).
  • .github/ — GitHub Copilot: hooks/, agents/, prompts/, copilot-instructions.md.
  • .agents/skills/ — cross-tool skills (Copilot, Zed, Codex).
  • AGENTS.md — this file; read automatically by Codex, Zed, and GitHub Copilot.
  • scripts/ — hook and workflow scripts (vs-record.sh, check-versions.sh, detect-manifest-edit.sh, detect-install-cmd.sh, auto-record.sh, prereq-check.sh).
  • commands/ — Claude Code slash commands (/vs-record, /check-versions).
  • platforms/kimi/commands/ — Kimi Code command files (same workflow, $KIMI_PLUGIN_ROOT-based).
  • skills/version-sentinel/ — agent skill explaining how to satisfy the hook.
  • hooks/hooks.json — hook definitions (Claude Code). Claude-only: the Bash handlers are gated with if rules that Codex ignores.
  • hooks/codex-hooks.json — the same wiring ungated, used by Codex.
  • tests/ — shell test suite; run tests/run.sh.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.