Replace line iter with bytes to lines wrapper

[sjmonson]

Summary

Read bytes stream rather then lines stream and manually split on double newline \n\n.

Details

Some models (gpt-oss) don't escape their newlines which causes httpx.iter_lines to split in the wrong spot. Note that is could potently could still be a problem if a model emits a \n\n token. so we may need a more robust solution in the future.

Test Plan

See original issue.

Related Issues

• Resolves orjson.JSONDecodeError (truncated JSON at char 158) when using /v1/completions with openai/gpt-oss-120b #662

---

• "I certify that all code in this PR is my own, except as noted below."

Use of AI

• Includes AI-assisted code completion
• Includes code generated by an AI application
• Includes AI-generated tests (NOTE: AI written tests should have a docstring that includes ## WRITTEN BY AI ##)

[jaredoconnell]

This does seem more like a workaround, but if this fixes the bug we can try it.

[sjmonson]

So this doesn't fully fix the issue but I think merge it anyways because it fixes a separate issue. We have been over-counting event iterations due to counting each newline when we really should have been counting each \n\n split. The actual gpt-oss issue may be unfixable. I thought this was working around a bug in httpx but in-fact the issue is that gpt-oss is emitting invalid JSON. The only way to fix it would be to manually correct the JSON which is either too computationally expensive or escapes valid JSON as well as invaild.

[dbutenhof]

This approach won't break anything worse than before, so if it avoids problems on even a subset of cases, it's a win. I agree: let's go for it.

[sjmonson]

Next release we can look at https://html.spec.whatwg.org/multipage/server-sent-events.html#event-stream-interpretation for building a better parser.