Yann AbadieHire me on Upwork

All selected work

Meta-YGN

A local safety runtime for AI coding agents. It reads what a shell command does, takes a checkpoint, and blocks destructive operations before they run.

Status
Status: Released
Stack
Rust, tree-sitter, SQLite FTS5, TypeScript, MCP
Last commit
19 November 2026
Sources
Each tag opens the README line that makes the claim, at commit ce4a595.

The problem

Coding agents run shell commands on real machines. One git reset --hard or find / -delete can wipe work nobody committed. A deny list of strings misses commands that do the same damage in other words: find / -delete contains no rm. README L38 (source in the Meta-YGN repository)

What I built

The README calls the runtime Aletheia-Nexus. README L1–3 (source in the Meta-YGN repository)

System as built

  1. Claude Code fires a hook

    On every tool call, before the command runs.README L114–123 (source in the Meta-YGN repository)

  2. AST Guard parses the command

    A tree-sitter syntax tree; effect and targets classified.README L37–38 (source in the Meta-YGN repository)

  3. Smart routing scores the risk

    In context: rm target/*.o scores 20, rm -rf / scores 0. Low means dangerous, the reverse of the illustration on the home page.README L40–41 (source in the Meta-YGN repository)

  4. Sequence monitor checks the chain

    Multi-step patterns such as clone, modify, force push.README L43–44 (source in the Meta-YGN repository)

  5. A second model, only if unclearonly when needed

    A Claude Haiku judge when AST analysis is inconclusive.README L46–47 (source in the Meta-YGN repository)

  6. Checkpoint before anything risky

    Git stash or file backup.README L49–50 (source in the Meta-YGN repository)

  7. Verdict back to the agent

    Allow, deny, ask or escalate, with recovery instructions.README L123 (source in the Meta-YGN repository)

How it is verified

What this means for your project

  • If your developers use coding agents, the same pattern puts a policy point between the agent and the shell: parse, score, checkpoint, then allow or ask. That is an Agent or MCP build.
  • The guard runs locally with no telemetry and no cloud calls, which suits teams with strict data rules. README L245 (source in the Meta-YGN repository)
  • An Architecture review can map which commands your agents run today and which should wait for a person.

Status & limits