Summary

codex_app_server currently appears to bypass Hermes' assembled prompt context for the main turn. In a profile/gateway session with a custom SOUL.md, the Codex subprocess receives the raw user input, but not the Hermes system prompt, persona, injected memory/context files, or current Hermes messages list.

That makes the runtime risky for persona-sensitive gateway profiles: storage/tools may still be profile-scoped via HERMES_HOME, but the model identity/voice is generic Codex rather than the configured Hermes profile.

Repro shape

1. Create/use a Hermes profile with a custom SOUL.md identity.

2. Enable:

   model:
     provider: openai-codex
     openai_runtime: codex_app_server

3. Run the profile through a gateway or CLI session.

4. Ask an identity/role question or short health check.

Expected

The Codex app-server turn should receive enough Hermes context to preserve the profile identity and normal Hermes behavior, or Hermes should clearly document/guard that this runtime is not persona-equivalent to the default runtime.

At minimum, one of these should be true:

• SOUL.md / assembled system prompt / memory/context files are seeded into the Codex thread.
• The current Hermes conversation history is projected into the Codex thread.
• The runtime is explicitly blocked or warned for gateway/persona-sensitive profiles.

Actual

The profile/gateway process is correctly scoped, and Codex can still access profile-local tools/storage through the Hermes MCP callback. However, the direct model behavior is generic Codex because the turn only sends the current user text.

This can create a split state:

• profile storage/tools: correct profile
• model identity/persona: generic Codex app-server context

Code path observed

• AIAgent.run_conversation() early-returns to _run_codex_app_server_turn() when api_mode == "codex_app_server".

• _run_codex_app_server_turn() accepts messages, but calls self._codex_session.run_turn(user_input=user_message).

• CodexAppServerSession.run_turn() sends:

  "input": [{"type": "text", "text": user_input}]

I did not see Hermes' assembled system_prompt, SOUL.md, injected memory/context, or messages being sent into the Codex thread on this path.

Impact

This is a UX regression for long-running messaging assistants whose core value is a stable profile identity and memory behavior. Users may think the wrong profile is attached even though HERMES_HOME and storage are correct.

Workaround

Use the default Hermes runtime for persona-sensitive profiles:

/codex-runtime auto

Keep codex_app_server for code-heavy sessions or experimental profiles where Codex's native shell/apply_patch/sandbox/plugin runtime is more important than Hermes persona continuity.