hermes - 💡(How to fix) Fix [Bug]: Codex Responses stream aborts on terminal null output [3 pull requests]

Bug

Hermes can abort openai-codex streams with TypeError: 'NoneType' object is not iterable after the backend closes the stream. The error is surfaced as non-retryable with HTTP None, suggesting it happens while the client processes streamed events rather than as an HTTP response.

Environment

• Provider: openai-codex
• Model: gpt-5.5
• Endpoint: https://chatgpt.com/backend-api/codex/responses
• OpenAI Python SDK observed locally: 2.24.0

Evidence

The failed request had a normal outgoing payload and no null values. Hermes had already created a Codex stream request. The stream then closed and Hermes raised TypeError: 'NoneType' object is not iterable.

Related existing reports/PRs found during triage include #11179, #32883, #11182, #32884, #32888, #32890, and #32891. This issue records the specific openai-codex/gpt-5.5 reproduction and recovery acceptance criteria from the handoff plan.

Likely Cause

The SDK Responses stream parser iterates response.output. If the terminal streamed response has output: null, the SDK raises before Hermes can use the streamed response.output_item.done events it already collected.

Expected Behavior

Hermes should recover from collected streamed output items or fall back to its raw create-stream parser before treating the failure as non-retryable.

Proposed Fix

Catch only the exact null-output parser TypeError in the Codex streaming path, synthesize the final response from collected output items when available, otherwise retry once and then use the existing create-stream fallback.