Skip to main content
This is the 0.5.0-beta.1 developer integration candidate tracked in #132. Publication and the installed-package acceptance result must be recorded there before treating this version as available. Use synthetic data while building the extension integration. The coordinated versions are Cargo/npm 0.5.0-beta.1 and Python 0.5.0b1. After publication, pin the exact package and commit its lockfile:
Node and WASM publish to npm’s beta dist-tag. The release must preserve latest.

​
Workspace handoff and restoration

Rust and WASM support persistent workspace pseudonyms, local preview and approval, strict all-or-nothing restoration, committed-marker follow-ups, atomic batches and encrypted backup/recovery through the reference host workflow. Damaged, unknown or unbracketed committed source markers block preparation before approval; safe byte, codepoint and UTF-16 ranges let a host highlight the input for editing. WASM releases retained stale plans when the saved head advances, while preserving valid current-head plans after no-op or failed activation. The host still owns credentials, encryption, durable storage and user approval. It must save the exact encrypted candidate with a complete-head comparison before calling activateSaved, and release text only after activation succeeds. A package upgrade does not add an extension vault, review UI or send-button enforcement. Follow the 0.5.0-beta.1 integration contract for Workspace, opaque plans, offset units, failure handling, follow-ups and close/reopen behavior. Its reference harness uses actual IndexedDB and local cryptography; test fault controls are not shipping extension features.

​
Optional English GLiNER2.5 Small in the browser

Import ModelSession from @datafog/wasm/models, explicitly provision the pinned model/tokenizer files (about 297 MB), and serve matching ONNX Runtime Web assets locally under the extension’s CSP. Open, scan and close the session explicitly. Pass the returned canonical findings into workspace preparation. Synchronous scan() remains rule-based. The browser model guide covers asset hashes, runtime paths and the unpacked MV3 proof. Requests are bounded to 65,536 UTF-8 bytes and 512 tokens including the schema. There is one inference at a time, with no inference cancellation/deadline or automatic chunking. English coverage and full-address ADDRESS spans require review; native GLiNER2 accuracy numbers do not qualify this browser model. Loading may stall its owning page and uses substantial memory; development-run measurements are not hardware guarantees. Native sessions also accept an explicit engine selector and expose configured engine/model/label metadata. Shared adapter construction, the model registry and Decima remain separate work; they do not block the current browser Small adapter.

​
Compatibility and acceptance

The workspace API and schema-1 snapshot format remain provisional. Pin the exact beta and recreate disposable development vaults on upgrades unless a migration has been explicitly qualified. This release does not promise real-data vault migration. The candidate is checked as a packed SDK in an unpacked Manifest V3 extension: model inference, offsets, approval and encrypted save before activation, failed-save non-release, strict restore, stale plans, missing/corrupt assets and browser restart. Final candidate hashes and local/CI evidence belong in #132. All SDK publication workflows also require the separate native model integration matrix for the exact release commit, followed by publication dry runs. Josh’s product still needs the extension integration and its own end-to-end acceptance against exact Core and extension commits. Finding sources and recall target 0.5.0-beta.2; later choices/rules and provenance/conversation work follow the versions or TBD assignments in the requirements spreadsheet. A real-data pilot additionally needs the snapshot/contract freeze, shipping-model qualification, provider marker tests and vault/backup review.