solopi-ai

Run doctor to verify device readiness, list cases with cases, then execute run --case <name> with a run timeout and artifacts directory. A run only counts as passed when the terminal state is passed, results are non-empty, and the exit code is 0.

Use verify-normalize, verify-compile, and verify-validate to turn requirements into a verified plan with a plan fingerprint, then execute it with verify-run. The Result Judge independently produces passed, failed, or not_tested based on checkpoint oracles and same-run evidence.

Yes, signed ExecuTorch model packages can be verified, installed, benchmarked, and activated through the model-* commands, then used via verify-run --decision-provider on-device. The model only proposes actions; the Result Judge remains the sole final verdict authority.

Yes, the managed-* commands provide a SQLite-backed device pool with persistent task queues, generation-leased assignments, and worker processes. The managed-report command emits stable CI exit codes with per-shard evidence digests.

case-validate only checks structure and parameter boundaries, not selectors, plugins, permissions, or business outcomes. Confirm the case fingerprint matches at launch, required plugins are installed and loaded, and verify with a real device replay to a terminal state.

The CLI deliberately offers no arbitrary shell, arbitrary scheme, or arbitrary HTTP mutation channels, and high-risk actions like clearData require explicit confirmation. System permissions, MediaProjection, and camera authorization must be confirmed by the user on the device.