feat: custom provider support, max tool turns, auto 429 retries, and …

[harshvishwakarma0]

Summary of Changes

This PR introduces comprehensive support for custom OpenAI-compatible providers (Ollama, LM Studio, Groq, Bynara, OpenRouter) and key user experience improvements around agent tool turn limits and error/retry resilience.

Key Features & Fixes:

1. Custom OpenAI-Compatible Provider Support:

   • Added full custom provider management in Settings UI with status indicators, model auto-fetching (/api/custom-providers/models), and local storage persistence.
   • Built CustomOpenAIProviderSession supporting Chat Completions streaming (client.chat.completions.create) and function calling serialization.

2. Customizable Max Agent Tool Turns:

   • Added an Agent Limits slider/numeric input in Settings (configurable from 5 to 50 turns).
   • Threaded maxToolTurns parameter through WebSocket pipeline and AgentEngine.

3. HTTP 400 Request Body & Rate Limit (429) Resilience:

   • Fixed content: null serialization issue in custom assistant messages that caused HTTP 400 errors on strict OpenAI proxies like Bynara router.
   • Added automatic RateLimitError (429) retry logic with exponential backoff in CustomOpenAIProviderSession.

4. Retry Data & Code Preservation:

   • Updated onVariantError and regenerate() in App.tsx so that partially-generated HTML code is preserved in version history when a run stops or errors out.
   • Clicking 'Retry' now sends an Update request containing existing generated code, resuming generation right where it left off instead of wiping project state.

Testing:

• Backend unit tests (poetry run pytest): All tests passing cleanly.
• Frontend TypeScript check (pnpm exec tsc --noEmit): 0 errors.
• Manual E2E verification with custom provider (Bynara / LiteLLM).

[abi]

@cursoragent review

[cursor]

You need to increase your spend limit or enable usage-based billing to run background agents. Go to Cursor

[abi]

@claude review