That is the whole design: one command answers from the artifacts the work left behind, and gating on its exit code means a false "done" cannot land. It reports what happened, not whether the code is good — quality stays with your tests and reviews.
A Python package you install into the git repo where your agents work. It reads the artifacts — git history, the file tree, the clock — and returns a verdict: SHIPPED, exit 0, or NOT_SHIPPED, exit 1. `dos-mcp` exposes the same verdicts as MCP tools.
- `dos_verify` — the agent checks its own last claim against what actually landed, and the exit code is the verdict
- `dos commit-audit` sweeps a commit range and flags the claims their own diffs do not back
- Collision detection when several agents touch the same files, and a flag when a run is spinning rather than progressing
- `dos init --hooks auto` detects the agent runtime already in the repo and wires the verdict into its hook config; `dos doctor` prints what the workspace is using
- Verdicts that travel onward — `dos export` drains the journal to your observability stack, `dos notify` pushes what needs a human, `dos attest` mints a signed receipt a skeptic can check
`pip install "dos-kernel[mcp]"` — the mcp extra is what installs the `dos-mcp` command — then add `{ "command": "dos-mcp" }` to your host config. Works on a plain git repo with no configuration. PyYAML is the only runtime dependency; CI covers Python 3.11–3.13 on Linux plus a Windows 3.13 smoke run.
One command — pip install "dos-kernel[mcp]"