Codex: move system prompts into user input for Responses API

[kumanday]

Summary

This PR fixes OpenHands Codex subscription model support by making the Responses API payload Codex-compatible. Codex subscription endpoints can reject complex/long top-level instructions (often returning 400 / “Instructions are not valid”), which breaks agent flows that rely on a large system prompt.

What changed

• Codex-specific message adapter for Responses formatting
  • For models whose id contains "codex":
    • We do not send any system messages as top-level instructions.
    • Instead, we merge system prompt content and prepend it to the first user input message (or inject a synthetic user message if none exists).
  • Non-Codex models keep the current behavior: system messages are concatenated into instructions.
• Ensure Codex models go through the Responses API path
  • Expanded Responses-model detection from only codex-mini-latest to any model containing codex (e.g. gpt-5.1-codex, gpt-5-codex, provider-prefixed variants).

Why this is needed

Codex subscription endpoints require server-validated instructions and can reject the complex, tool-heavy system prompts typically used by agent frameworks. Moving that content into user input avoids the strict validation while preserving the same guidance/context for the model.

Tests

• Added a unit test that verifies for a Codex model:
  • instructions is None
  • system prompt content appears at the start of the first user input message.

[all-hands-bot]

Found several issues that need attention: overly broad model matching, potential data loss bug, and missing test coverage. Details in inline comments.

[enyst]

I think this is actually the right idea, though I’m not sure about the details of the messages.

It seems codex wants its own system prompt basically, more or less