Add OpenAI support (gpt-4o, gpt-4o-mini, gpt-3.5-turbo)

[sajosam]

Summary

SpawnVerse currently ships with Groq/LLaMA as the default provider.
This issue adds first-class OpenAI support so users can run any
SpawnVerse task on GPT-4o, GPT-4o-mini, or GPT-3.5-turbo with a
single config change and zero modifications to their task code.

What is already done

The multi-provider _ProviderClient wrapper is merged into core/engine.py.
The _complete_openai() method exists. The agent stdlib think() function
already has the OpenAI branch. Auto-detection of gpt-* model names works.

What still needs doing

• Integration test _complete_openai() on all three models
• Confirm resp.usage.total_tokens tracking works as expected
• Verify rate-limit retry (HTTP 429 → exponential backoff)
• Uncomment openai>=1.30.0 in requirements.txt
• Uncomment the OpenAI block in examples/providers/run.py
• Update README model table with GPT rows

Acceptance criteria

• All three GPT models complete a full SpawnVerse run
• Token budget is enforced correctly across agents
• HTTP 429 triggers the retry path and recovers

Skills needed

Python · OpenAI SDK · basic pytest

[IlyasFardaouix]

I think the missing piece here is verifying the rate-limit retry (HTTP 429 → exponential backoff) for the OpenAI provider. Have you tried adding some logging or print statements to the _complete_openai() method to see if it's actually triggering the retry logic when encountering a 429 response? This should help us confirm whether the issue is with the retry mechanism itself or with the OpenAI API being rate-limited.

[sajosam]

@IlyasFardaouix

The retry logic lives in the central _llm() function in engine.py, not inside _complete_openai() directly — so both providers share the same backoff loop. The check is if "429" in err or "rate_limit" in err.lower(). If you want to test it, temporarily lower token_budget to force exhaustion and watch the logs. Happy to add explicit retry logging if that helps debugging.

[IlyasFardaouix]

@sajosam, I agreethat the retry logic is in the _llm() function. To further debug this, I suggest we temporarily lower the token_budget in the examples/providers/openai.py file to force exhaustion and verify if the retry logic kicks in. We can then inspect the logs to see if the exponential backoff is being applied correctly.

[sajosam]

@IlyasFardaouix
Actually those are two separate code paths in _llm() — lowering token_budget hits the budget guard at the very top and returns "", 0 immediately, so the retry loop never even runs. You won't see a 429 that way.
To actually test the retry branch: set "rate_limit_wait": 1 in config and drop a quick print(f"attempt={attempt} waiting={wait}s") inside the backoff loop, then crank "max_parallel": 8 and let it hammer the API until you hit a 429. That'll confirm the exponential backoff is firing correctly.