SDK & Tools

Release notes for the sovseal Python SDK, TypeScript SDK, stdio Model Context Protocol (MCP) server, and developer CLI tools.

Authorization Header Fix: Resolved a critical issue where remote MCP servers (cloudflare and Google stitch) were missing their required authorization and API key headers in args, which caused the entire MCP client startup to block or hang while waiting for interactive browser OAuth.
Enabled Core MCP Servers: Enabled the Google stitch and TestSprite MCP servers by default to ensure all automated workflows and tests can access their respective tools.

Type Checker Fixes: Resolved strict compilation errors across the @sovseal/mcp-server package, improving the developer experience.
Zod Schema Resolution: Hardened Zod schema typing to seamlessly support custom schemas with default or optional parameters where input and output types differ.
Apache Arrow Compatibility: Added explicit type declarations to resolve compiler warnings for users importing the Apache Arrow module alongside the SDK.

Sub-5ms Recall: Drastically reduced semantic recall latency. Hot query times now reliably hit ~4.2ms.
LRU Query Caching: Implemented a 256-capacity LRU cache for query embeddings to eliminate duplicate ONNX model calls (configurable via SOVSEAL_EMBEDDING_CACHE_SIZE).
Eager Pipeline Warmup: The MCP server now fires a background no-op embedding immediately upon connection, eliminating the 1-3 second ONNX cold-load delay on the agent's first memory query.

On-Device Embedding Pipeline: The SDK now integrates Transformers.js for on-device embedding generation (auto-downloaded ~30MB model on first run) and LanceDB for local vector storage. Your agent's context never leaves your machine unless explicitly synced.

Removed Key-Value Tools: The exact-match save_context and load_context tools have been completely removed and replaced with semantic tools: store_memory and recall_memory (in @sovseal/[email protected] / Platform v3).
Conceptual Querying: The new semantic tools embed memories locally and rank them by L2 distance, allowing the agent to query conceptually (e.g., "what did we agree about X?") instead of guessing exact string keys.
Migration path: Simply update your agent prompts to use store_memory({ content }) and recall_memory({ query, topK }). No API keys or network calls are required.

Recall Validation Primitives: On every state restore, the client now derives a SHA-256 hash of the decrypted payload and compares it against the stored server hash. If there is a mismatch (due to data corruption or tampering), the system fails closed, providing mathematical proof of state integrity.