fix: build a chat-completions client for v0 eval in the legacy bridge

[mikasenghaas]

Summary

Fixes v0 env eval crashing in the legacy bridge.

The orchestrator drives both training and eval rollouts of a v0 env through the same LegacyEnvServer (run_rollout / run_group → _run_v0 → _v0_client). Training uses a renderer (token-in/out) client; eval uses an OpenAI chat-completions client (pool.get_eval_client()). The guard added in #1613 raised on any non-renderer config:

ValueError: the v0 legacy bridge trains through a renderer (token-in/out) client,
got client type 'openai'; MITO (chat-completions) training of v0 envs is not supported

so it broke v0 eval.

_v0_client now dispatches on the config type instead of raising:

• RendererClientConfig (type="renderers", training) → v0 renderer client (unchanged).
• OpenAIClientConfig (type="openai", eval) → v0 chat-completions client.

Verification

• ruff check --isolated / ruff format --isolated --check clean.
• _v0_client(OpenAIClientConfig(), model) builds an OpenAIChatCompletionsClient (no longer raises); a RendererClientConfig builds the renderer client as before.

Note

Support OpenAI chat-completions client config in LegacyEnvServer._v0_client

Previously, _v0_client in legacy.py raised a ValueError for any non-renderer client config. It now accepts OpenAI chat-completions configs and builds a V0ClientConfig with client_type="openai_chat_completions", omitting renderer-specific fields. Renderer configs continue to pin the tokenizer to renderer_model_name as before.

^{Macroscope summarized 06b45e2.}

---

Note

Low Risk
Narrow adapter change in the legacy bridge; renderer training path is unchanged and eval now mirrors existing _eval_client behavior.

Overview
Fixes v0 environment eval when rollouts go through LegacyEnvServer (orchestrator eval uses OpenAIClientConfig, not the renderer used for training).

LegacyEnvServer._v0_client no longer raises on non-renderer v1 configs. It branches on config type: RendererClientConfig still maps to a v0 renderer client (tokenizer pinned via renderer_model_name); any other config (eval’s OpenAI chat-completions) maps to v0 openai_chat_completions with base URL, API key var, and headers only. Client caching is unchanged.

Training-only rejection of chat-completions for v0 MITO is removed in favor of supporting eval while keeping the renderer path for training.

^{Reviewed by Cursor Bugbot for commit 06b45e2. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR removes an explicit error guard and enables a previously-blocked code path for chat-completions clients in the v0 legacy bridge. While the change is small, it enables new runtime behavior rather than fixing broken functionality, warranting human review.

^{You can customize Macroscope's approvability policy. Learn more.}