atrislabs Engines
An engine is the intelligence that runs a bounded Atris task. The orchestrator owns scope and verification. Engines do not self-certify.
Roster
Show installed engines and the current default:
atris engine
Switch the workspace default:
atris engine <name>
atris engine reset
Run one command on a specific engine without changing the default:
atris mission run "<objective>" --engine <name>
atris autopilot --auto --iterations=1 --engine <name>
atris run --engine <name>
Preflight an engine when available:
atris engine test <name>
Engine Choices
atris-fast: cheap, fast, bounded lookups and small edits.codex: deep root-cause work and longer autonomous builds.cursor: fast repo-local edits and refactors.devin: multi-step feature work, especially when the run should survive the local machine.claude: broad reasoning, critique, and implementation when installed as a CLI engine.
Use the local roster output as truth. If an engine is not ready, do not route work to it until the binary or login is fixed.
Dispatch Contract
Every engine task needs:
- Absolute repo path.
- One bounded objective.
- Files or systems in scope.
- Exit criteria.
- Verify command.
- Git rules: status first, stage only own files, never destructive git.
- Final report: files changed, verifier result, residual risk.
Verification
After an engine returns, independently inspect the diff and rerun the verifier. If the engine wrote task receipts inside a sandbox or worktree, reconcile them with the live Atris task or mission state.
Rules
- One writing engine per checkout.
- Use worktrees for parallel or risky work.
- Treat engine output as a claim until verified.
- Never let the builder be the only certifier.