Your Coding Agent’s External Conscience

What it does

Vet snapshots a repository and its diff, optionally ingests the agent’s conversation history, and runs LLM checks to find logic errors, security flaws, and mismatches between what you requested and what the agent produced. It filters and deduplicates the results into a final list of issues with predictable exit codes.

The interesting bit

Unlike static analyzers, Vet inspects the agent’s intent by reading conversation logs, flagging cases where the agent misunderstood the goal or wandered off-task. It also installs as a “skill” into Claude Code, Codex, and OpenCode, forcing the agent to self-critique immediately after editing.

Key highlights

• Bring-your-own-model via OpenAI-compatible endpoints, Anthropic, OpenAI, or Gemini, plus a community model registry.
• Runs as a terminal CLI, a reusable GitHub Action for PR review, or an embedded agent skill that triggers automatically.
• Customizable issue guides via TOML to tune detection for specific bug classes like SQL injection or integer overflow.
• Predictable exit codes—exit 10 specifically means issues were found, making CI integration straightforward.

Caveats

• The --history-loader option executes arbitrary shell commands as the current user to retrieve conversation history; the README explicitly warns that you must audit these commands before use.

Verdict

Teams using AI coding agents who want an external sanity check on both diff correctness and goal adherence will get the most value. If your workflow is already locked down by traditional linters and you don’t delegate to agents, this is likely unnecessary.