hermes - 💡(How to fix) Fix Codex Responses adapter can send invalid function_call.name from replayed history

Bug Description

Long-running Hermes sessions using the openai-codex / Codex Responses adapter can fail with an HTTP 400 from the backend when replayed conversation history contains a function_call.name that does not match the stricter Responses API name pattern.

Observed error:

BadRequestError: HTTP 400
Invalid 'input[222].name': string does not match pattern.
Expected a string that matches the pattern '^[a-zA-Z0-9_-]+$'

This blocks GPT-5.5 / Codex-backed Hermes sessions even though the configured tool schemas themselves are valid.

Why this happens

The adapter validates that function_call.name exists, but it only applies .strip() before sending replayed conversation items to the Responses API.

Relevant paths in agent/codex_responses_adapter.py:

• _chat_messages_to_responses_input(...): emits replayed assistant tool_calls as Responses function_call items
• _preflight_codex_input_items(...): normalizes pre-built Responses input items
• _preflight_codex_api_kwargs(...): normalizes Responses tools
• _responses_tools(...): converts chat-completions tool schemas

The Responses backend requires every emitted name to match:

^[a-zA-Z0-9_-]+$

A name with spaces, dots, or other punctuation in replayed history can therefore trigger a backend 400, even if all current tool definitions are valid.

Expected Behavior

Hermes should sanitize or reject invalid Responses function/tool names before sending the request, so the user gets either:

1. a valid request, or
2. a clear local validation error identifying the offending item.

Actual Behavior

The invalid name is sent through to the Codex Responses backend, which rejects the entire request with a 400 such as:

Invalid 'input[222].name': string does not match pattern

Suggested Fix

Add a small sanitizer for Responses names and apply it wherever names are emitted into Responses API payloads:

_RESPONSIVE_NAME_RE = re.compile(r"[^A-Za-z0-9_-]")


def _sanitize_responses_name(name: str) -> str:
    return _RESPONSIVE_NAME_RE.sub("", name)

Then use it for:

• function call names from chat-history tool calls
• function call names in preflighted Responses input items
• tool definition names in Responses tool schemas

A local patch using this approach compiled cleanly and preserved normal names like browser_click and web_search unchanged, while converting examples like some.name with spaces to somenamewithspaces.

Environment

• Hermes repo: NousResearch/hermes-agent
• Provider route: openai-codex
• Transport: Codex Responses adapter
• Model route affected: GPT-5.5 / Codex-backed Responses API
• OS: Linux

Related

Possibly related to #11411, but this variant is specifically about non-empty names with invalid characters in replayed Responses input[n].name items.