Changelog
Generated from CHANGELOG.md in the project repository. Never maintained twice.
Unreleased
- addedOpenAI-compatible proxy:
POST /v1/chat/completionswith streaming passthrough, andGET /healthz. - addedToken estimation over message history and tool schemas, rule-based classification into five task types, and a privacy scan that forces local-only routing.
- addedTier-based selector: quality floor, capability, context and provider filters, budget-pressure downgrade, cheapest-viable choice and a remaining-budget gate.
- addedRetry with jittered backoff, error classification and fallback along the candidate chain.
- addedExact-match SQLite cache with TTL, history guard and eviction; optional semantic cache.
- addedOptional LLM classification by a small local model, with fallback to the rules.
- addedNative Anthropic adapter, weighted multi-key providers, model aliases and parameter sanitisation.
- addedRouting governance: project scopes, routing profiles, latency as a routing and reporting dimension, and a capability catalog.
- addedRequest pipeline as an ordered plugin chain for budget, cache and logging.
- addedSQLite request log with policy result per request, and a read-only Stats API (
/api/stats/*, eight endpoints). - addedAstro dashboard, embedded in the binary and served at
/. - added
Dockerfileanddocker-compose.yml. - addedConfig lookup in the user directory,
llmux init, andllmux --demowith a built-in echo provider. - addedPer-request override headers
x-llmux-*. - securityConstant-time comparison of the gateway key.