What does this PR do?

Fixes a custom-endpoint compatibility bug in the Codex Responses streaming path.

Some provider: custom + api_mode: codex_responses relays emit a valid, usable SSE stream but start with response.in_progress instead of response.created. curl -N can still read the stream, but the OpenAI SDK's strict Responses stream state machine raises:

Expected to have received response.created before response.in_progress

This PR keeps the strict responses.stream(...) path as the default, but adds a narrow compatibility fallback for custom providers only:

• retry once on the normal stream path
• if the same missing-response.created ordering error persists, fall back to Hermes' existing responses.create(stream=True) parser
• tolerate fallback streams that begin with response.in_progress
• preserve live text/reasoning deltas and output backfill while using the fallback

The behavior for native Codex / non-custom providers stays unchanged.

Related Issue

Fixes #8133

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• Scoped missing-response.created detection to custom Codex Responses providers in run_agent.py
• Reused the existing responses.create(stream=True) fallback instead of broadening the main SDK path for every provider
• Extended the fallback parser to handle streams that start with response.in_progress
• Added fallback-side delta/reasoning callback delivery so live streaming still works during compatibility fallback
• Added regression tests for:
  • custom provider retries + fallback when response.created is missing
  • non-custom provider still raising on the same SDK error
  • fallback parser consuming a stream that begins with response.in_progress

How to Test

1. Configure Hermes with a custom Responses endpoint:

   model:
     provider: custom
     default: gpt-5.4
     base_url: https://<custom-endpoint>/openai
     api_key: <key>
     api_mode: codex_responses

2. Reproduce against a relay that omits the initial response.created event and starts its SSE stream with response.in_progress.
3. Confirm Hermes no longer fails with Expected to have received response.created before response.in_progress and instead completes normally.
4. Run the targeted regression tests:

   .venv/bin/pytest tests/run_agent/test_run_agent_codex_responses.py -q
   .venv/bin/pytest tests/run_agent/test_streaming.py -q

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 26.4

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Screenshots / Logs

Observed failing SDK error before this patch:

Expected to have received response.created before response.in_progress