diff --git a/packages/llm-llamacpp/CHANGELOG.md b/packages/llm-llamacpp/CHANGELOG.md index 7dd622482a..1006db6834 100644 --- a/packages/llm-llamacpp/CHANGELOG.md +++ b/packages/llm-llamacpp/CHANGELOG.md @@ -1,5 +1,34 @@ # Changelog +## [0.45.0] - 2026-08-19 + +This release removes sliding-context support. The addon no longer evicts tokens +from the middle of the KV cache when the context fills, so a full context is +reported to the caller instead of being made room for. + +### Breaking Changes + +- The `n_discarded` load option is no longer supported. Configurations that pass + it now fail model loading as an unsupported option; remove the key. +- The `contextSlides` runtime stat is gone from `runtimeStats()` and from the + `RuntimeStats` type in `index.d.ts`. +- A conversation that outgrows the context no longer keeps going by dropping + old tokens. Behaviour is now what `n_discarded=0` already did, which was the + default: + - a prefill that does not fit throws `ContextOverflow`, carrying the cached + token count, the prompt token count and the context size; + - a generation that fills the window stops with + `stopReason: "contextOverflow"` and still returns the tokens it produced. +- `ContextSlideFailed` is retired as an error code. Its numeric value stays + reserved so the codes after it do not shift; the one remaining internal + failure it covered now reports `FailedToDecode`. + +### Changed + +- Per-session cache metadata keeps its four-field width, so cache files written + by this version still load on older builds and vice versa. The two + first-message fields are retired and written as `0`. + ## [0.44.0] - 2026-08-17 ### Added diff --git a/packages/llm-llamacpp/package.json b/packages/llm-llamacpp/package.json index 830882d7a6..d0720628ec 100644 --- a/packages/llm-llamacpp/package.json +++ b/packages/llm-llamacpp/package.json @@ -1,6 +1,6 @@ { "name": "@qvac/llm-llamacpp", - "version": "0.44.0", + "version": "0.45.0", "description": "llama addon for qvac", "addon": true, "scripts": {