Repository navigation
Releases
Adaptive Reasoning 0.2.3 — portable defaults and history recovery
Adaptive Reasoning 0.2.3 is suitable for fresh installations: routes start empty, preserving the chosen model's default reasoning until an explicit supported provider/model mapping is configured.
Includes the 0.2.2 history-replay correction: routing diagnostics use a separate sidecar rather than adding unsupported events to DSH conversation history. Existing custom routes are not replaced by this package's defaults.
Validation: 63 unit tests; real DSH integration and compressed cold-history reopening; structured-speech greeting fixtures for off/xhigh; packed installation/removal proof. See README and docs/SETUP.md for installation and configuration.
Adaptive Reasoning 0.2.0 preview
First downloadable preview of Adaptive Reasoning 0.2.0. Runtime policy matches the locally installed 0.2.0; public setup instructions and package metadata were added.
One-time model configuration is required outside the tested Augmentor/Qwen setup. The stock allowlist is intentionally narrow and is inactive for other presets and models. Follow Setup.
The plugin uses local rules to select reasoning per request. It adds no classifier call, GPU keep-alive or recurring job. Unknown/consequential work keeps high effort. Prompt improvement remains high because faster candidates failed the quality checks.
Download dsh-adaptive-reasoning-0.2.0.tgz. For the exact tested route:
dsh plugin --profile web add https://github.com/ManoloRemiddi/dsh-adaptive-reasoning/releases/download/v0.2.0/dsh-adaptive-reasoning-0.2.0.tgz --ignore-scripts --config.auto-install-peers=falseFinish active tasks and restart DSH after installation. This preview was verified against DSH 0.1.5-rc.1 and Cordis 4.0.2. It is not a universal difficulty estimator or a measured electricity-saving guarantee.
Validation: 48 tests, actual DSH/adapter integration against a deterministic local endpoint, and disposable packaged install/composition/removal passed. See the README and validation docs for limitations and earlier finite model tests. The SHA256SUMS asset verifies the download.