Bug Description

PR #12846 (which closed #8294) enabled Anthropic prompt caching for third-party gateways using api_mode: anthropic_messages, but only when the model name contains "claude". This leaves out providers that use the native Anthropic protocol for their own model families.

MiniMax is a concrete example. It serves MiniMax-M2.7, MiniMax-M2.5, etc. through https://api.minimax.io/anthropic (and api.minimaxi.com/anthropic for China). MiniMax documents explicit cache_control support on this endpoint:

• Docs: https://platform.minimax.io/docs/api-reference/anthropic-api-compatible-cache
• Supported models: M2.7, M2.5, M2.1, M2
• Cache read pricing: 0.1x input; write: 1.25x input; TTL: 5 min

Currently, _anthropic_prompt_cache_policy() returns (False, False) for MiniMax because is_claude is False for model="minimax-m2.7".

Expected Behavior

MiniMax (both built-in minimax / minimax-cn providers and custom endpoints pointing at MiniMax's Anthropic URLs) should receive cache_control breakpoints with the native layout, just like native Anthropic and Claude-on-third-party gateways.

Actual Behavior

No caching. The system prompt + tool schemas are re-billed in full on every turn.

Why this is a Hermes-side issue

MiniMax's Anthropic-compatible endpoint is already handled correctly for auth (Bearer token, stripped betas) and transport (anthropic_messages). The only missing piece is the cache-policy gate in _anthropic_prompt_cache_policy().

Scope beyond MiniMax

This same pattern likely affects any future provider that:

1. Speaks the native Anthropic protocol (anthropic_messages)
2. Documents cache_control support
3. Serves its own (non-Claude) model family

A capability-based or provider-allowlist approach would be more future-proof than the current is_claude name check for third-party gateways.

Version

Current upstream main (post #12846).