Hermes Mistral Cache Affinity

What Hermes Mistral Cache Affinity does

The plugin acts only when a request is in chat_completions mode, goes over HTTPS to api.mistral.ai or api.eu.mistral.ai, and carries a non-empty prompt_cache_key. It then adds x-affinity with that key to extra_headers. It does not change messages, tools, the system prompt or the cache key, and it returns a copy rather than mutating the original request. An explicit x-affinity header in any letter case is preserved, and for other providers, hosts, schemes or API modes it does nothing.

The README reports that Mistral sessions plateaued around 25 percent cache hit before the plugin and reached 93 to 95 percent in observed sessions, citing verification/live-report.json and agent.log cache lines. It describes the change as a routing improvement, not a guarantee, since a prefix change, expiry or provider behavior can still cause a miss. The plugin performs no network access, logging or secret handling of its own. Tests include synthetic unit tests, integration tests that need a hermes-agent checkout, and an opt-in live test against the Mistral API.

Key features

• Adds x-affinity from the existing prompt_cache_key
• Limited to HTTPS Chat Completions requests to the two Mistral hosts
• Leaves messages, tools, system prompt and cache key unchanged
• Preserves an explicit x-affinity header
• Unit, integration and opt-in live tests with privacy-preserving reports

When to use it

• Improving prompt cache hit rates when Hermes talks to Mistral
• Cutting repeated prefill cost in long Hermes sessions on Mistral models
• Choosing the EU or global Mistral endpoint through the provider base_url

Who it is for: Hermes Agent users who run Mistral models and want better prompt cache reuse.

How it fits with Hermes Agent

It is a Hermes plugin that hooks the llm_request middleware and is enabled per profile in config.yaml.

Note: The README presents it as a cache routing improvement, not a cache-hit guarantee, and the repository has no license file.

FAQ