Fix: propagate top-level llm.provider to per-model configs

[froucoux]

Summary

A top-level llm.provider (e.g. claude_code) set in a config was silently dropped and never reached the individual models, so the ensemble always fell back to the OpenAI backend. This makes the examples/claude_code_quickstart example crash immediately with:

openai.OpenAIError: Missing credentials. Please pass an `api_key` ... or set the `OPENAI_API_KEY` ... environment variable.

even though config.yaml correctly sets provider: "claude_code".

Root cause

LLMConfig propagates shared settings down to each per-model LLMModelConfig via a shared_config dict (in both __post_init__ and rebuild_models). That dict included api_base, api_key, temperature, etc. — but not provider. So each model kept provider=None, and LLMEnsemble._create_model routed it to OpenAILLM instead of ClaudeCodeLLM.

>>> c = Config.from_yaml("examples/claude_code_quickstart/config.yaml")
>>> c.llm.provider
'claude_code'
>>> [m.provider for m in c.llm.models]
[None, None]          # <-- provider never propagated

Fix

Add "provider": self.provider to both shared_config dicts. Propagation goes through update_model_params(..., overwrite=False), so an explicit per-model provider still takes precedence over the top-level default.

Tests

Added TestProviderPropagation to tests/test_claude_code_llm.py:

• top-level provider reaches every evolution and evaluator model,
• an explicit per-model provider overrides the top-level one,
• the default provider stays None (unchanged OpenAI-default behavior),
• LLMEnsemble built from such a config yields ClaudeCodeLLM instances end to end.

The three propagation tests fail on main (None != 'claude_code') and pass with this change. Full suite: 417 passed (unchanged), plus the new cases.

How it was found

Running the documented examples/claude_code_quickstart example with the Claude Code CLI backend on Python 3.14. After the fix, a full 50-iteration run completed successfully, evolving the seed random-search into a simulated-annealing optimizer (combined_score 1.2148 → 1.4995).

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codelion]

Thanks for this — the diagnosis was spot on, and the reproduction made it easy to confirm:

>>> c = Config.from_dict({"llm": {"provider": "claude_code", "models": [{"name": "a"}]}})
>>> [m.provider for m in c.llm.models]
[None]          # provider never propagated

Since the CLA is unsigned we could not merge this branch directly, so the fix was reimplemented independently and has shipped in v0.3.1 via #457 (merged as 8bd33ad):

• provider added to both shared_config dicts (__post_init__ and rebuild_models), so a top-level llm.provider now reaches every evolution and evaluator model. Propagation uses overwrite=False, so an explicit per-model provider still wins.
• New tests/test_provider_propagation.py covering propagation to both model lists, per-model override precedence, the unchanged None default, survival across rebuild_models(), and end-to-end that LLMEnsemble now actually constructs ClaudeCodeLLM rather than silently falling back to OpenAILLM.

Closing as superseded — but the credit for finding and diagnosing this is yours. If you do sign the CLA at some point, we would be glad to take future contributions directly.