Skip to content

v0.80.0

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 06:09
· 20 commits to main since this release

What's Changed

  • fix(identity): replace keystore files in place on Windows by @Virgile-pct in #2986
  • test(runtime): skip layered package parts when picking a cached GGUF by @Virgile-pct in #2988
  • test(xtask): expect the workspace version in the package-source-version contract by @Virgile-pct in #3006
  • fix(plugins): keep plugin files inside their root by @benthecarman in #2913
  • docs(windows): make the quickstart work after setup started the background service by @Virgile-pct in #3000
  • feat(setup): say what to run next at the end of setup by @Virgile-pct in #3004
  • feat(routing): route cache hits on the current request's cost against cold candidates by @Virgile-pct in #2951
  • docs: use the meshllm.cloud install URL everywhere by @Virgile-pct in #3014
  • docs(quickstart): end the Windows service to recover from a failed load by @Virgile-pct in #3013
  • ci(windows): catch release packaging failures in PR and main CI by @michaelneale in #2641
  • fix(ci): admit recovered retry packages at Llama Upgrade publication by @i386 in #3002
  • docs(plugins): point to the Mesh-LLM example plugin by @i386 in #3022
  • docs(website): add Extend entry point and released plugin directory by @michaelneale in #3023
  • test(skippy): quiet test transcripts behind a SKIPPY_LOG_LEVEL gate by @i386 in #3001
  • fix(website): generate CLI inventory on native Windows by @Bortlesboat in #2989
  • fix(cli): give the largest subcommands their own Args structs so parsing needs less stack by @kleinbem in #3017
  • fix(windows): replace the node key, config and private files in place by @Virgile-pct in #3007
  • fix(discovery): prefer public listings whose publisher serves for thin clients by @Virgile-pct in #2983
  • feat(routing): assess which candidates meet a first-token latency target by @Virgile-pct in #2985
  • test(skippy-ffi): update llama_context_params layout for b12199 by @Virgile-pct in #2987
  • fix(stop): end the Windows background service instead of a node it restarts by @Virgile-pct in #3005
  • test(skippy): test the default DFlash selection on Windows too by @kleinbem in #2969
  • fix(log-store): drop the needless return clippy rejects on Windows by @kleinbem in #2964
  • fix(skippy): re-resolve a layer package from the hub when its cached snapshot is schema v1 by @Virgile-pct in #2967
  • docs(ci): record the admitted Python SDK checkpoint and its ABI prerequisite by @ndizazzo in #2974
  • fix(skippy): enforce reasoning budgets in chats by @benthecarman in #3020
  • feat(routing): fair single-model mesh/auto routing with MoA as an opt-in plugin by @michaelneale in #2856
  • fix: skip unsupported model providers in Auto and mesh by @michaelneale in #3024
  • ci: remove unused Docker workflow and pause nightly runs by @i386 in #2896
  • test(plugin): turn installed plugins off explicitly in the exchange conformance harness by @michaelneale in #2959
  • feat(plugins): let plugins rewrite inference requests at ingress (request_transform.v1) by @michaelneale in #3012
  • feat(runtime): count speculative verify-window tokens in each stage's meter by @AwokenLazarus in #2949
  • fix(llama): certify upstream a4d4f2fd4e by @i386 in #2992
  • fix(runtime-events): point the prefill family at its producer after #2856 by @Virgile-pct in #3036
  • perf(ci): stop running the Claude Code test serially in PR Linux planning by @ndizazzo in #3033
  • fix(build): package the Windows runtime with Git Bash, not WSL bash by @Virgile-pct in #3019
  • perf(ci): build a narrow automation binary for protected authority producers by @ndizazzo in #3034
  • test(skippy): give the late-worker-reply deadline test time to enqueue by @kleinbem in #3018
  • perf(ci): build a narrow runner-cleanup binary for the trusted runner image job by @ndizazzo in #3039
  • fix(ci): pin Windows sccache to 0.16.0 so Rust 1.100 builds fit the command line by @ndizazzo in #3047
  • test(process-supervision): run the sleep the test process finds on PATH by @kleinbem in #2939
  • fix(console): remove unreliable catalog architecture labels by @michaelneale in #3029
  • feat(mesh): let nodes refuse peer model downloads by @benthecarman in #3048
  • fix(runtime): keep a cached Hugging Face GGUF's requested identity as its declared ref by @i386 in #3051
  • fix(runtime): publish thinking observations for startup-served models by @i386 in #3052
  • fix(llama): certify upstream a4d4f2fd4e by @i386 in #3049
  • feat(routing): @private/@Incognito requests that stay on your own listed nodes by @michaelneale in #3037
  • docs: organize roadmap by focus area by @matbalez in #3032
  • fix(ci): publish selected Llama candidates after recovered retries by @i386 in #3053
  • fix(cli): load the installed native runtime before gpus, models and download by @Virgile-pct in #3043
  • docs: list Mesh maintainers by @michaelneale in #3056
  • fix(ci): limit Mac Llama native tests to four processes by @i386 in #3061
  • perf(ci): run cheap Quality consistency checks before the contract suite by @ndizazzo in #3060
  • fix(discovery): replace dead and paid default Nostr relays by @michaelneale in #3063
  • fix(release): bind packaging dispatch to release identity by @ndizazzo in #3071
  • ci: refresh admitted runner image cohort by @ndizazzo in #3068
  • perf(ci): start each product when its own runtime is ready by @ndizazzo in #3069
  • fix(release): await verified packaging readiness by @ndizazzo in #3073
  • feat(private): set up private routing with mesh-llm private add/list/remove by @michaelneale in #3058
  • fix(ci): keep lane summaries aligned with frozen workflows by @i386 in #3083
  • fix(cli): keep main's clippy green after the private and runtime-load merges by @michaelneale in #3085
  • feat(analytics): add model, hardware, workload and speculation dimensions to serving events by @danielwinterw in #3082
  • feat(mesh): maintain the Python SDK in the main repository by @i386 in #3072
  • feat(sdk): transcribe embedded audio in process by @i386 in #3064
  • feat(sdk): expose structured inference errors across SDKs by @i386 in #3062
  • feat(sdk): expose local runtime capacity by @i386 in #3059
  • fix(llama): certify upstream 8954f19483 by @i386 in #3080
  • fix: preserve response identity across native chat streams by @i386 in #3088
  • fix(analytics): report the hardware profile after the native runtime loads by @Virgile-pct in #3075
  • feat(windows): embed the MeshLLM icon in mesh-llm.exe by @Virgile-pct in #3076
  • fix(skippy-serving): continue a prefilled assistant message instead of closing it by @Virgile-pct in #2997
  • perf(ci): reuse verified automation across SDK and smoke jobs by @ndizazzo in #3089
  • fix(mesh): show one public mesh under one mesh ID by @michaelneale in #3087
  • fix(gossip): stop answering reconnect loops with the full peer table by @ethanyzhang in #3078
  • chore(release): prepare 0.80.0 version metadata by @michaelneale in #3103

Full Changelog: v0.79.0...v0.80.0