Agent: LM Studio as a local model; remove the Claude Code mode

[drthyang]

Summary

LM Studio as a fourth way to reach a model. LM Studio 0.4.1+ answers the Anthropic Messages API at /v1/messages, so the Agent drives it with the same chat loop, tools, approval cards and undo as Ollama. It sends only plain headers and only the request fields a local server reads.

• The drawer lists the server's language models from /api/v1/models, those trained for tool use first, each with the context its loaded copy was given.
• It auto-picks a tool-trained model with room for the Agent's ~15k-token instructions, and warns when a model is loaded with less than 32k (lms load <model> --context-length 32768).
• Errors name the CORS switch (LM Studio admits no other origin until Enable CORS is on), the version requirement, a context overflow, and a server that requires a token.

The Claude Code mode is removed, together with the materia-live MCP bridge (src/agent/bridge/, bridgeClient.ts), scripts/live-bridge-demo.mjs, the mcp:live / build:mcp-live scripts and its .mcp.json entry.

• The materia MCP server is unchanged.
• The Agent now starts on the user's own API key; a saved Claude Code choice falls back to it, and stored settings keep only known keys.
• With one sender left, tool calls no longer carry a source.

Shared local-server helpers (plainFetch, the 32k context floor) move to src/agent/localServer.ts.

Docs: docs/AGENT_TOOLS.md (modes table, LM Studio, a plain explanation of the local proxy), docs/ROADMAP.md, CHANGELOG.md (Unreleased).

Testing

• npm run typecheck, npm run lint, npm run build, and node scripts/build-mcp.mjs all pass.
• Full suite: 1674 passed, 113 skipped.
• New tests:
  • lmstudio.test.ts: model listing against a stand-in server in LM Studio's documented shape, plus the too-old and not-LM-Studio cases.
  • chat.test.ts: LM Studio request headers and fields, the Ollama → LM Studio history switch, and error messages.
  • useAgent.test.ts: migrating stored settings from Claude Code.
• Ran the dev app in Chromium against a stand-in LM Studio (CORS on):
  • a stored claude-code setting migrated to API key;
  • the models listed and the auto-pick skipped the 4k-context one;
  • a chat turn called get_state and answered, sending no x-api-key, anthropic-* or authorization headers;
  • a wrong address showed the CORS hint.
• Not tested against a real LM Studio install (not downloadable from this environment).

🤖 Generated with Claude Code

https://claude.ai/code/session_01MReUfYLPkA6mqDmVQK5NpM

---

Generated by Claude Code