openenv collect --llm-endpoint rejects any URL that includes a port or path #1188

Summary

--llm-endpoint is documented as an "OpenAI-compatible endpoint URL", but the value is handed to OpenAIClient(endpoint, port), which builds its base URL as f"{endpoint}:{port}/v1". Any URL that already names a port is mangled, and any path prefix gets /v1 appended after it.

$ openenv collect reasoning_gym:chain_sum --base-url http://localhost:8001 \
    --output-dir ./rollouts --llm-endpoint http://localhost:8000 --model Qwen/Qwen3.5-4B -n 2
Error: Invalid port: '8000:8000'

The only working form is the undocumented split --llm-endpoint http://localhost --llm-port 8000. It is shown in examples/ttt_collect_with_llm.py but not in --help or anywhere under docs/.

Where it lives

• src/openenv/core/llm_client.py — LLMClient.base_url is f"{self.endpoint}:{self.port}"; OpenAIClient.__init__ passes f"{self.base_url}/v1" to AsyncOpenAI.
• src/openenv/cli/commands/collect.py — --llm-endpoint help text says "endpoint URL"; --llm-port defaults to 8000.

The same constructor backs LLMJudge (core/rubrics/llm_judge.py) and build_model_step (core/harness/collect.py), so the limitation applies to every self-hosted judge or teacher, not only the CLI. Servers behind a path prefix (a LiteLLM proxy, vLLM behind a reverse proxy, Azure's /openai/v1) cannot be reached at all.

Repro without a server

from openenv.core.llm_client import OpenAIClient
OpenAIClient("http://localhost:8000", 8000, model="x")
# httpx.InvalidURL: Invalid port: '8000:8000'
OpenAIClient("http://proxy:4000/litellm", 4000, model="x")._client.base_url
# http://proxy:4000/litellm:4000/v1/

Proposed fix

Treat endpoint as a real URL in LLMClient: append /v1 only when the URL has no path and use a URL with a path as-is (the OpenAI SDK convention, so gateway prefixes such as Gemini's /v1beta/openai keep working), append port only when the URL names none, reject an explicit port that contradicts the URL, and raise a clear ValueError for malformed endpoints. Existing ("http://localhost", 8000) callers are unchanged; --llm-endpoint http://localhost:8000 and http://localhost:8000/v1 become equivalent. --llm-port should default to unset so a full URL without a port is left alone. Add a collect section to the CLI reference while at it, since the command is not documented there.

I have this working end to end against vLLM on a local GPU and will open a PR.

Environment: main @ 0521700, Python 3.12, openai 3.14.1.