feat: v2 - ai assistant llm proxy and dispatcher agent

[maxy-shpfy]

Description

Wires the @openai/agents SDK into the in-browser AI assistant, replacing the previous echo-only worker stub with a real LLM-backed dispatcher.

A new TangleDispatcher wraps an @openai/agents Agent instance configured to use a user-supplied LLM proxy token stored in IndexedDB via localforage. Static configuration (proxy base URL, model names, proxy mode) is read from a committed src/config/aiAssistantConfig.json file. The proxy token is resolved per-turn from IndexedDB and passed into ensureProxyConfigured(), which rebuilds the OpenAI client only when the token rotates.

The worker's init(onStatus) / ask() lifecycle is preserved. createWorkerApi() now instantiates the dispatcher on init() and delegates ask() calls through it, resolving the token from IndexedDB, creating a per-thread MemorySession, and mapping the result back to the existing AgentResponse shape.

Observability hooks translate @openai/agents lifecycle events (agent_start, agent_end, agent_tool_start, agent_handoff) into human-readable status strings forwarded to the main thread via the Comlink-proxied StatusCallback.

A AiTokenSetup component is shown in the chat panel when no token is stored, allowing the user to paste and save a proxy token without leaving the editor. The token is persisted to IndexedDB and the React Query cache is updated immediately on save.

The Vite worker bundle is configured with a scoped resolveId plugin that forces @openai/agents-core/_shims to resolve to its browser variant, working around Vite's worker bundler not reliably applying the browser export condition. A globalThis.process stub in the worker entry point covers an unguarded process.env read in @openai/agents-core v0.4.x.

Related Issue and Pull requests

Type of Change

• Bug fix
• New feature
• Improvement
• Cleanup/Refactor
• Breaking change
• Documentation update

Checklist

• I have tested this does not break current pipelines / runs functionality
• I have tested the changes on staging

Screenshots (if applicable)

AI Assistant - Set token and first LLM.mov (uploaded via Graphite)

Test Instructions

1. Set proxyBaseUrl and orchestratorModel in src/config/aiAssistantConfig.json.
2. Open the editor and navigate to the AI chat panel.
3. Paste a valid LLM proxy token into the token setup form and click Save token.
4. Send a message. Verify the status line updates during the turn (e.g. "Thinking...", "Preparing response...").
5. Verify the assistant responds with a relevant answer and the thread ID is stable across follow-up messages.
6. Reload the page and confirm the token is persisted and the setup screen is not shown again.
7. Clear the token from IndexedDB (e.g. via DevTools) and confirm the setup screen reappears.

Additional Comments

src/config/aiAssistantConfig.json is committed with empty strings for proxyBaseUrl and orchestratorModel. These must be filled in before the assistant will function. The subagentModel field is reserved for a future sub-agent PR and is not read yet.

[github-actions]

🎩 Preview

A preview build has been created at: 05-26-feat_v2_-_ai_assistant_llm_proxy_and_dispatcher_agent/65cce3b

[maxy-shpfy]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• chore: v2 - ai assistant - unit tests #2336
• feat: v2 - ai assistant - architect subagent #2335
• refactor: v2 - ai assistant - split tools bridges #2333
• feat: v2 - ai assistent - debug agent #2332
• feat: v2 - ai assistent - repair subagent #2331
• feat: v2 - ai assistant - tangle help subagent #2330
• feat: v2 - ai assistant llm proxy and dispatcher agent #2329 👈 (View in Graphite)
• feat: v2 - ai assistant web worker scaffolding #2328
• feat: v2 - ai assistant chat window #2327
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.