Summary

Using Hermes with model.provider=openai-codex, first turn succeeds but second turn fails consistently with Responses input errors related to replayed reasoning items.

Environment

• Hermes: v1.0.0
• OpenAI SDK: 2.24.0
• Codex CLI: 0.106.0
• OS: Ubuntu 24.04.4 LTS
• Python:
  • Hermes venv: 3.11.14
  • system python: 3.12.3
• Config:
  • model.default: gpt-5.3-codex
  • model.provider: openai-codex
  • terminal.backend: docker
  • TERMINAL_ENV=docker

Repro

1. Configure Hermes for openai-codex.
2. Start a chat.
3. Send first prompt (hello) -> success.
4. Send second prompt -> failure.

Observed errors

1. ValueError: Codex Responses input[1] has unsupported item shape (type='reasoning', role=None).
2. 400 invalid_request_error: Missing required parameter: 'input[1].summary'.

Request payload detail

In a failing dump, replayed input contains:

{
"type": "reasoning",
"encrypted_content": "..."
}

(no summary field)

Suspected root cause

• Hermes replays prior reasoning items for Codex continuity.
• Preflight may reject type=reasoning replay items.
• Backend appears to require summary on replayed reasoning items.

Local patch/workaround that resolved it

• Accept replayed reasoning items in preflight.
• Preserve/include summary when replaying reasoning.
• Add fallback summary: [] for older persisted sessions with only encrypted_content.
• Drop malformed reasoning items gracefully instead of hard failing.

Validation

• Targeted tests pass after patch:
  • python -m pytest -q tests/test_provider_parity.py -k 'reasoning or preflight'
  • 10 passed

Request

Please confirm expected replay schema for Codex reasoning items (summary required?) and align Hermes replay/preflight contract with backend expectations.