Highlights
- Model selection is now active-safe. Saving
/modelduring running work leaves every admitted turn and native steering operation on its captured model, while the first later provider dispatch—including a queued continuation—uses the newly committed model. - Communication prompts now carry a framework-owned evidence-grounded honesty contract even when a workspace preserves a customized chat template. It requires explicit separation of observation, evidence, inference, uncertainty, and unknowns, with proportionate authoritative checks for material claims.
Compatibility and migration
No configuration or durable-state migration is required. Harness implementation and sandbox changes remain idle-only; only forward-looking model selection gained the active-safe boundary. Existing in-flight work is deliberately not restarted, relabeled, or mutated.
Spynel remains alpha software and supports the current schemas only. The prompt contract materially strengthens expected behavior but is not a mathematical guarantee that a model cannot make mistakes.
Platforms and packaging
Native archives are published for Linux amd64/arm64 and macOS amd64/arm64, with target-matched sherpa-onnx and ONNX Runtime libraries and required notices. Windows remains temporarily unsupported by both native packaging and the npm launcher. Model weights remain checksum-pinned first-use downloads rather than release assets.
Validation
The frozen source passed focused race-enabled coverage for model dispatch ordering and the prior release-critical job-history, cancellation, and notification paths, plus the full Go test/vet/race suite, build, DOX validation, smoke test, npm launcher tests, npm pack dry-run, diff hygiene, and a runnable Linux arm64 native archive smoke.