Xenon v0.7.3 — Cache Rails
This release makes prompt-cache continuity a first-class session invariant while preserving model-routing flexibility.
Highlights
- Per-model append-only Cache Rails survive model switches and keep separate engine, phase, tool-contract, and context-epoch boundaries.
- Working memory, retrieved memory, per-turn guidance, and execution boundaries are frozen into immutable request envelopes instead of rewriting earlier prompt positions.
/cache lanes, Prompt Manifest, Cache Doctor, history, and cost reporting expose lane generations, forks, reusable-token estimates, and provider-reported hit/miss usage without storing prompt text.- AutoRouter may use a recent exact-prefix lane only as a weak tie-breaker between otherwise suitable healthy models; it never overrides explicit selection or capability constraints and sends no prewarming requests.
- DeepSeek model identities and cache telemetry are normalized across direct streaming and engine calls.
Real-provider validation
- Flash / Pro alternated for 10 calls, 5 calls per model.
- Two independent active lanes, zero lane forks.
- Across two complete runs, all 16/16 warm returns reported provider cache hits.
- Warm input-token hit rates: 96.57% and 96.74%.
Quality gates
1662 passed, 38 deselectedin the full offline suite.- Ruff, compileall, and
git diff --checkpassed. - Wheel and sdist builds passed Twine metadata validation.
Install
pip install -U "git+https://github.com/xianyu-sheng/Xenon.git@v0.7.3"
xenon --versionSee CHANGELOG.md and the DeepSeek cache guide for details.