v2.11.2 v0.76.0
cheaper-inference/OmniRouteSep 9, 2026 → Sep 10, 20260 days between releases
What's New
Major architectural overhaul introduces durable KV cache and radix architecture, replacing flat caches and persisting state, while introducing several breaking behavior changes.
Breaking Changes
- Stop persisting KV cache state to disk
- Remove stale disk cache defaults
- Consolidate model=auto and model=mesh routing
- Treat max_tokens as a ceiling, not a context reservation
- Stop aborting healthy long prefills at five minutes
New Features
- Durable KV prefix cache that survives restarts
- Unified radix cache architecture replacing flat caches
- Effective configuration schema enforcement
- Durable lifecycle logging and audit APIs
- Iteration-level scheduler for concurrent staged serving
Improvements
- Reuse warm prefixes across agent sessions
- Route repeated prompts to workers with verified radix cache reuse
- Stream tool calls generation as they are produced
- Support staged Qwen3.8 Flash-Next inference
- Complete graph-derived v2 model splitting
Bug Fixes
- Deduplicate reference-coverage records to prevent fatal errors
- Avoid false parser limits for simple runtime parameters
- Account for distinct reference targets on the same source line
- Count unquoted characters separately to prevent false command-word span limits
- Preserve partial coverage for printf and wrapper paths
Upgrade Notes
The new release replaces flat prefix caches with a unified radix architecture and introduces durable KV caching. Configuration settings previously inert are now enforced by a schema, and max_tokens is now treated as a ceiling rather than a context reservation.
The Two Releases
SkillSpector v2.11.2
Sep 9, 2026· 1 month ago
Fixes fatal reference accounting errors and false static-parser limits triggered by documentation, preserving incomplete analysis reporting for runtime-selected executables.
Full release notesv0.76.0
Sep 10, 2026· 1 month ago
This major release introduces a durable KV prefix cache that survives restarts, replaces flat prefix caches with a unified radix architecture, and makes previously inert configuration settings effective. It also significantly enhances observability with durable lifecycle logging and audit APIs.
Full release notes