Docs
Everything, in the order you need it.
Start with install and the quickstart. The concept pages explain why the gate is absolute and how routing decides what to spend.
Getting started
Guides
- Fixing a real bugWriting an issue that carries evidence, choosing the target test, and what to do with a failure.
- Bring your own providerAny OpenAI-compatible endpoint, your own key, and the settings file that saves you retyping both.
- Running across repositoriesFan a task set out through the scheduler, one supervised agent per bug.
- The dashboardA read-only web view over logs a run already wrote.
Concepts
- The verifier gateWhy success is a test result rather than a claim.
- Adaptive routingPredicting per-call difficulty, and paying for the expensive model only when it is needed.
- The sandboxWhat the agent can reach, and what it cannot.
- The four layersHarness, execution, runtime and memory — what each owns, and the contracts that hold them together.
- Decision memoryA decision store and a code graph the planner reads before it plans, and the ablation that measured what that changed.
Reference
- CLI referenceEvery command, with the entry point resolved against pyproject.toml.
- MCP toolsFive tools over stdio, plus a client for consuming external servers.
- ConfigurationThe two-tier settings layout, the real precedence chain, and the task config keys the harness and runtime read.
- Trace eventsThe per-task trace.jsonl — every prompt, response, tool call and result, one JSON object per line.