mcp sampling

[pelikhan]

---

Pull Request Summary

MCP Client Sampling Integration 🚀

• Introduced MCP Client Sampling Support:

  • Added a new language model provider "mcp" that enables sampling via the Model Context Protocol (MCP) client.
  • Implemented mcpRequestSample to handle chat completion requests through the MCP server and return responses using a streaming protocol.

• Worker Language Model Bridge 🧑‍💻

  • Created a createWorkerLanguageModel utility, allowing worker threads to delegate language model requests back to the parent process for sampling.
  • Enhanced worker and API logic to support passing chat completion requests and responses between processes.

• API and Server Enhancements 🔗

  • Extended the run API to support an onMessage callback with a postMessage function for bidirectional communication.
  • Enabled the server to detect and use client sampling capabilities, forwarding chat completion requests to the client when "mcp" is the selected model.
  • Updated server initialization to check client capabilities and set up the appropriate language model bridge.

• Resource Management Improvements 🗂️

  • Fixed a typo in resource upsert method (upsetResource → upsertResource).
  • Ensured resource updates are handled correctly via event listeners and upsert logic.

• LLM Data and Model Resolution Updates 🧠

  • Registered "mcp" as a valid model provider and updated model resolution logic to support it.
  • Added metadata and configuration for the "mcp" provider in the LLMs data registry.

• Sample Script Update 📝

  • Modified the emojifier sample script to use the new "mcp" model provider for demonstration.

• Type and Code Quality Improvements 🧹

  • Improved type annotations and consistency across multiple modules.
  • Refactored event handling and watcher logic for better reliability and maintainability.

---

These changes collectively enable MCP-based client sampling for language model completions, allowing for advanced, pluggable model usage and improved extensibility in the platform.

AI-generated content by prd may be incorrect.