fix: preserve gpt-5.5 streamed text when response.completed.output is empty

[sungwon1110]

Summary

gpt-5.5 (and possibly other newer Codex backend models) emits response.completed with an empty output: [] array even though text is streamed via response.output_text.delta events. The current parser preferentially returns response.completed.response and discards accumulated deltas, so downstream callers (Claude Code via this proxy, etc.) receive an empty content block.

Older models (gpt-5.4, gpt-5.3-codex) populate output[] in the completion event, so the bug only surfaces on newer models.

Symptoms

• Claude Code through this proxy with model=gpt-5.5 consistently received empty assistant text.
• Claude Code looped on retry / context-compact because responses appeared blank.
• tool_diag log line showed codex_text_blocks=0 anthropic_text_blocks=1 — one block emitted, but the text inside was empty.
• Reproducible with a one-line curl:

  curl -s -X POST http://127.0.0.1:19080/v1/messages \
    -H 'x-api-key: dummy' -H 'anthropic-version: 2023-06-01' \
    -H 'Content-Type: application/json' \
    -d '{"model":"gpt-5.5","max_tokens":50,"stream":false,
         "messages":[{"role":"user","content":"Reply: alpha"}]}'
  # → {"content":[{"type":"text","text":""}]}   (before fix)
  # → {"content":[{"type":"text","text":"alpha"}]}   (after fix)

Root cause

The response.completed event from the Codex backend looks like this for gpt-5.5:

{"type":"response.completed","response":{
  "model":"gpt-5.5",
  "output":[],
  "usage":{"output_tokens":28,"output_tokens_details":{"reasoning_tokens":21}}
}}

Note output: [] even though response.output_text.delta events arrived earlier with the actual text.

Fix

Apply the same defensive merge in both code paths inside src/codex/client.ts:

• parseSseResponse (used when client requests stream:true)
• parseFinalResponse (used when client requests stream:false; the proxy still streams to upstream and collects, so the same shape arrives here)

Both paths now:

1. Accumulate response.output_text.delta deltas alongside capturing finalResponse.
2. After the stream ends, check whether finalResponse.output contains any non-empty output_text block.
3. If not, append a synthetic message block with the joined deltas so downstream extractContentFromCodexOutput sees the text.

Older models still hit the original code path (since their output[] is non-empty); behavior unchanged.

Test plan

• Manual curl against /v1/messages with model=gpt-5.5 (both stream:true and stream:false) returns the expected text.
• Same call with model=gpt-5.4 still returns text (regression check on the populated-output path).
• Claude Code session against the patched proxy stops looping on compact and produces visible responses.
• Add a unit test (happy path + empty-output-with-deltas) — happy to follow up if maintainers prefer.

🤖 Generated with Claude Code