The kernel that doesn't believe your AI agents.
DOS (dos-kernel) catches AI coding agents when they lie about what
they shipped — by reading the evidence they couldn't have written.
An AI agent will tell you it finished. DOS checks the real world instead of taking its word — and the nearest piece of the real world is your git history. Run a dozen agents on one repo and DOS tells you which claim of “done” is real, which agents are stepping on each other's files, and which run is spinning in circles. Every verdict comes from artifacts (git, the file tree, the clock), never from the agent's narration.
Try it in 60 seconds
pip install dos-kernel
dos quickstart # stages an agent lying about a commit — and catches itdos verify AUTH AUTH1 # SHIPPED AUTH AUTH1 e62f74d (exit 0)
dos verify AUTH AUTH2 # NOT_SHIPPED AUTH AUTH2 (exit 1)Works on a plain git repo with zero config. One small Python package; the only runtime dependency is PyYAML. MIT licensed.
What it answers
dos verify— did this work actually ship? Answered from git ancestry, never from the agent's self-report.dos arbitrate— may this agent take these files right now? A deterministic referee between parallel agents; no double-booking, structured refusals.dos liveness— is this run still moving, or just spinning? Judged from the git and journal delta, not from “making progress…”.dos top/dos decisions— live terminal views of what's running now and what needs a human.
Works where your agents work
A plain CLI with exit-code verdicts (drop it into CI), an
MCP server
exposing the verdicts as tools, and a
Claude Code plugin
that wires the checks in as hooks. Host integrations cover Claude Code, Codex,
Gemini CLI, Cursor, and more via dos init --hooks.