Bug Description

When using third-party Anthropic-compatible providers (Kimi, MiniMax, DashScope, etc.) via api_mode: anthropic_messages, the agent incorrectly self-identifies as "Claude Code" and replaces all mentions of "Hermes Agent" with "Claude Code" in the system prompt.

Root Cause

_is_oauth_token() in agent/anthropic_adapter.py only excludes sk-ant-api* prefixed keys. Any third-party API key (e.g. sk-kimi-*) returns True, causing _is_anthropic_oauth to be set, which triggers:

1. System prompt prepend: "You are Claude Code, Anthropic's official CLI for Claude."
2. Text replacement: "Hermes Agent" → "Claude Code", "Nous Research" → "Anthropic"
3. Tool name prefixing with mcp_

Additionally, this code path hits two secondary crashes:

• AttributeError: 'NoneType' object has no attribute 'get' on self.request_overrides.get("speed") when gateway sets request_overrides = None
• UnboundLocalError on api_kwargs in the retry-exhaustion error handler

Steps to Reproduce

1. Configure a Kimi API key (sk-kimi-*) with provider: custom and api_mode: anthropic_messages
2. Send a message via the gateway (Telegram/Discord/Slack)
3. Ask the agent "who are you?"
4. Agent responds as "Claude Code" instead of "Hermes Agent"

Expected Behavior

Third-party providers using the Anthropic Messages API should not get Claude Code identity injection.

Environment

• Hermes Agent v0.8.0
• Provider: Kimi (api_mode: anthropic_messages)
• Platform: Telegram gateway