opencodex 2.100.2 passes Qwen2.5-32B API translation test

The test focused on a practical pain point for developers who use coding agents: each agent tends to expect a specific API format, while local models served through Ollama speak OpenAI-compatible endpoints. opencodex addresses this by translating requests on the fly, and the writer confirmed that translation works not just for plain text but for streaming responses and the tool-call round trips that coding agents rely on. The absence of measurable proxy latency matters because the model itself generated at only about 6 tok/s on the test hardware, leaving little room for conversion overhead to hide.

The writer also highlighted operational caveats. The official README warns that some providers, notably Anthropic, may suspend or restrict accounts that route traffic through third-party proxies, so the test used curl instead of subscription logins. The tool's startup also launches a proxy that intercepts traffic to api.anthropic.com and includes code that writes to Claude Code settings files, which is why the writer isolated the home directory during testing. A separate bug surfaced in the Anthropic-format response: the model field was replaced with "the requested model" because the validation regex rejects the colon in Ollama model names, though the response content and tool calls were unaffected.