[Bug] Responses-family parsers drop the terminal SSE event and unparented text deltas

[Adolanium]

Affected area

AI providers and models

What happened?

Two silent data-loss bugs in the OpenAI Responses-family parsers.

1. Codex parseSSE only yields events delimited by a blank line inside the read loop. If the server closes right after the last data: line, the terminal response.completed event is lost, along with usage and stop reason. CRLF is also not tolerated.
2. The shared Responses handler drops output_text.delta and refusal.delta when no content_part.added arrived first. Azure, OpenRouter, and proxies do this, and the streamed text vanishes with no error.

Steps to reproduce

1. Serve a Codex SSE stream that ends with a final data: line and no blank line, then close the connection. The client never sees response.completed.
2. Stream response.output_text.delta without a preceding response.content_part.added. The assistant text is gone.

Expected behavior

Flush remaining SSE after the read loop, same as the Anthropic iterator, and normalize CRLF. Lazily create the missing content part before appending a delta.

Prime Agent version

0.7.1 / source as of 2026-08-08

Environment

Reproduced from source. The delta drop shows up on Azure, OpenRouter, and other Responses-compatible endpoints that skip content_part.added.

Additional context

Original report: #995

Tested patch, closed in the August 15 backlog sweep: #996

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: Codex parseSSE failed to parse the trailing output stream when no newlines delimited terminal events, losing the response.completed terminal event. Additionally, Responses-family handlers in openai-responses-shared would silently drop stream deltas like output_text.delta if a prior content_part.added event was skipped (a behavior observed on some third-party endpoints).

Fix: Modified parseSSE to handle CRLF line endings, and flush its buffers on conclusion. Modified openai-responses-shared to lazily create output_text or refusal parts when deltas arrive while the parts array is empty.

Validation: I wrote and executed regression tests for both issues (test/suite/regressions/codex-parse-sse-trailing.test.ts and test/suite/regressions/responses-missing-content-part.test.ts). Ran npm run check and vitest, verifying fixes passed reliably.

Upstream discussion: #1770

Diff: awhite0030/prime-agent@main...fix-responses-parser-data-loss-15600850845829273638
Full pull request: awhite0030#13

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.

[Adolanium]

@awhite0030 Same situation as on #1769. I diagnosed this in #995 and shipped the patch in #996 on Aug 8. This discussion restates both bugs and both fixes. Your fork PR implements them.

A few things that aren't independent rediscovery:

• Same two edits: normalize CRLF and flush leftover SSE after the read loop, and lazily create the missing content part on output_text.delta / refusal.delta.
• Test file is mine with the prefix dropped: openai-responses-missing-content-part.test.ts vs responses-missing-content-part.test.ts.

The comment frames this as you investigating the report and producing a candidate fix. If this goes anywhere, credit #995 / #996 and @Adolanium.

Please don't do this on other people's reports either. If a discussion already points at an issue and a closed PR, credit that work. Don't present the same fix as yours.