Bug description

When an agent turn reaches agent.max_turns while running a codex_responses model through a named custom/fallback provider, Hermes tries to generate the iteration-limit summary but crashes with:

I reached the maximum iterations (60) but couldn't summarize. Error: 'CodexAuxiliaryClient' object has no attribute 'responses'

The original agent turn correctly hit the iteration limit, but the fallback summary path should still produce a summary. Instead, the summary path receives a CodexAuxiliaryClient wrapper where it expects a raw OpenAI client exposing .responses.

Environment

• Hermes Agent: v0.12.0 (2026.4.30)
• Commit: 75e1339d4
• Python: 3.11.15
• OpenAI SDK: 2.32.0
• OS: macOS-26.4.1-arm64-arm-64bit
• Runtime: Discord gateway session

Relevant configuration shape

Sanitized config excerpt:

model:
  default: gpt-5.5
  provider: custom-or-named-provider
  api_mode: codex_responses
  context_length: 256000

custom_providers:
  - name: named_custom_responses_provider
    base_url: https://example.invalid/openai/v1
    key_env: EXAMPLE_API_KEY
    api_mode: codex_responses
    model: gpt-5.5

fallback_providers:
  - provider: named_custom_responses_provider
    model: gpt-5.5
    base_url: https://example.invalid/openai/v1
    key_env: EXAMPLE_API_KEY
    api_mode: codex_responses

agent:
  max_turns: 60

Steps to reproduce

1. Configure a named custom provider with api_mode: codex_responses.
2. Configure it as a fallback provider, or otherwise route a long-running gateway/CLI turn through that named custom provider.
3. Run a task that exhausts agent.max_turns.
4. Hermes enters the iteration-limit summary path.
5. Summary generation fails with CodexAuxiliaryClient missing .responses.

Actual behavior

Gateway logs show the provider fallback and then summary failure:

INFO root: Fallback activated: gpt-5.5 → gpt-5.5 (named_custom_responses_provider)
WARNING root: Failed to get summary response: 'CodexAuxiliaryClient' object has no attribute 'responses'
INFO gateway.run: response ready: ... api_calls=60 response=127 chars

User-visible response:

I reached the maximum iterations (60) but couldn't summarize. Error: 'CodexAuxiliaryClient' object has no attribute 'responses'

Expected behavior

When a codex_responses turn reaches the iteration limit, Hermes should generate a summary using a client that supports the Responses API (client.responses.create/stream) or otherwise route through the correct transport adapter. It should not surface an internal wrapper attribute error to the user.

Likely root cause

In agent/auxiliary_client.py, the named custom_providers branch appears to wrap the OpenAI client with CodexAuxiliaryClient whenever the custom entry declares api_mode: codex_responses:

if entry_api_mode == "codex_responses" and not isinstance(client, CodexAuxiliaryClient):
    client = CodexAuxiliaryClient(client, final_model)

However, some main-agent paths call resolve_provider_client(..., raw_codex=True) because they need direct client.responses.* access. CodexAuxiliaryClient only exposes client.chat.completions.create(...) compatibility and does not expose .responses, so the main-agent Responses path crashes when it receives the wrapped client.

This differs from the _needs_codex_wrap() helper, which already respects raw_codex:

if raw_codex:
    return False

Suggested fix

In the named custom provider branch, respect raw_codex before wrapping:

if (
    entry_api_mode == "codex_responses"
    and not raw_codex
    and not isinstance(client, CodexAuxiliaryClient)
):
    client = CodexAuxiliaryClient(client, final_model)
else:
    client = _wrap_if_needed(client, final_model, raw_base_for_wrap, custom_key)

or route this branch through _wrap_if_needed() consistently so the existing raw_codex guard is honored.

Additional note

Raising agent.max_turns only delays the first condition. It does not fix the .responses attribute error in the iteration-limit summary path.