Skip to content

v0.79.0

Choose a tag to compare

@github-actions github-actions released this 10 Nov 22:01
· 107 commits to main since this release

[0.79.0] - 2026-11-10

Added

skippy

plugins

mesh

console

runtime

  • Add runtime.native_runtime.download to start without runtime downloads by @kleinbem in #2496
  • Warn when a detected GPU is left unused for lack of a fitting runtime by @kleinbem in #2528
  • Load native models on demand and unload idle instances by @i386 in #2704
  • Count decode steps on every split stage and report each stage's meter by @AwokenLazarus in #2902

sdk

Other

Changed

  • Use one MeshNode across platform SDKs by @i386 in #2241
  • Turn the speculation gate on by default by @danielwinterw in #2283
  • Name shared inference API contracts by protocol scope by @i386 in #2229
  • Harden serving against hostile requests and peers by @benthecarman in #2307
  • Shorten decode stalls from hybrid-model cache checkpoints by @benthecarman in #2393
  • Avoid discarded prefill buffer copies by @micspiral in #2396
  • Resolve a model ref against one cached repository by @i386 in #2424
  • Build the CUDA runtime with CUDA graphs again by @Virgile-pct in #2573
  • Accelerate grammar sampling with the upstream sampler by @i386 in #2677
  • Serve document OCR under /mesh/v1 by @i386 in #2708
  • Hash each native runtime file once per process by @Virgile-pct in #2694
  • Add mesh-wide discovery and consistent model APIs by @i386 in #2703
  • Make the node payment-agnostic with generic plugin seams any payment plugin can use by @michaelneale in #2654
  • Stop reading the whole executable at startup when it has no release footer by @kleinbem in #2751
  • Roll back split verify windows in place by @danielwinterw in #2736
  • Send split DFlash feature rows only inside the draft window by @danielwinterw in #2906
  • Send split DFlash feature rows as Q4 blocks when the final stage reads them by @danielwinterw in #2910
  • Prefill splits in 512-token chunks that may span several frames by @danielwinterw in #2923
  • Keep prefix-cache copies and the coordinator tick off the decode path by @danielwinterw in #2937
  • Stop bundling plugins with MeshLLM releases by @i386 in #2977

Fixed

skippy

  • Probe all accepted reasoning efforts by @i386 in #2274
  • Show native model-open error text in logs by @i386 in #2277
  • Report a finished serving task instead of racing the API probe by @danielwinterw in #2284
  • Make the speculation gate able to catch its own motivating case by @danielwinterw in #2288
  • Fail the fixture generator when a fixture is not written by @Virgile-pct in #2280
  • Initialize replayed recurrent inputs with an empty slot by @Virgile-pct in #2259
  • Align Decisions text contract with OpenAI docs by @i386 in #2304
  • GPU benchmark fails when the NVIDIA driver is older than the CUDA toolkit by @kleinbem in #2303
  • Size the ggml scheduler split context for the graph being split by @Virgile-pct in #2330
  • Derive KV payloads from admitted graph state by @i386 in #2252
  • Prefer a local native runtime over an equally ranked download by @Virgile-pct in #2413
  • Runtimes are rejected when the system glibc differs from the one mesh-llm runs on by @kleinbem in #2438
  • Build model splits from live graphs by @i386 in #2414
  • Repair canary classification and vision verification by @i386 in #2460
  • Release a stalled prediction return stream on Windows by @Virgile-pct in #2455
  • Read the local model cache without a CA bundle by @kleinbem in #2498
  • Detect AMD GPU architectures on Windows before choosing ROCm by @Virgile-pct in #2567
  • Serve a local model without a CA bundle by @kleinbem in #2491
  • Avoid inventory scans during L3 spill admission by @i386 in #2667
  • Fix split serving for MTP packages, deep splits and coordinator takeover by @danielwinterw in #2734
  • Cap native MTP to one draft token on hybrid and recurrent models by @ndizazzo in #2658
  • Decode temperature-0 requests with ignore_eos or logit bias at greedy speed by @ndizazzo in #2660
  • Skip prefix-cache snapshots that would not fit in host memory by @ndizazzo in #2680
  • Repair main after the bindgen and DFlash split-load merge by @i386 in #2815
  • Restore split admission for derived-persistent stage state by @danielwinterw in #2911

runtime

  • Serve external endpoints without a native runtime by @ndizazzo in #2225
  • Keep the --discover name when a join token picks the mesh by @Virgile-pct in #2553
  • Name the Skippy runtime release in the missing-runtime error and stop redacting " / " by @kleinbem in #2527
  • Bind public listeners exclusively on Windows by @Virgile-pct in #2679
  • Ship the rocBLAS and hipBLASLt kernels in the Windows ROCm runtime by @Virgile-pct in #2722
  • Require in-process worker readiness for embedded startup by @ndizazzo in #2643
  • Require owner-control credential for owner-control routes by @ndizazzo in #2645
  • Keep native decode timing on successful generation terminals by @ndizazzo in #2655
  • Apply request_defaults config changes to running models by @ethanyzhang in #2760
  • Advertise only Windows ROCm targets whose rocBLAS kernels are packaged by @Virgile-pct in #2797
  • Don't report a Vulkan ICD beside CUDA or ROCm as a second unused GPU by @kleinbem in #2698
  • Run the runtime and doctor commands without a CA bundle by @kleinbem in #2577
  • Say a CUDA GPU's architecture is unknown when none was detected by @kleinbem in #2868
  • Give up auto lanes before overrunning with an explicit DFlash draft by @Virgile-pct in #2888

mesh

  • Allow Auto tools requests when capability metadata is missing by @michaelneale in #2239
  • Reject split plans that exceed lane-scaled KV capacity by @i386 in #2275
  • Choose usable models before automatic routing by @michaelneale in #2557
  • Renew the node ownership certificate while the node runs by @Virgile-pct in #2569
  • Retry a failed save of the renewed ownership certificate by @Virgile-pct in #2610
  • Set context and lanes per load and report slow or failed loads clearly by @ndizazzo in #2656
  • Restore GGUF size/arch metadata lost since the skippy split by @michaelneale in #2651
  • Stop a peer from pinning unlimited plugin stream handlers by @ndizazzo in #2930
  • Preserve HF resolve URL identity and bound host frame reads by @ndizazzo in #2957

payments

cli

  • Derive pseudo-subcommand value-taking flags from clap by @iamthesvn in #2400
  • Warn on all bare-surface flags, stop --discover swallowing subcommands by @iamthesvn in #2422
  • Show the dashboard for bare automatic mode by @michaelneale in #2535
  • Define the top-level options in groups so parsing fits a 2 MiB stack by @kleinbem in #2690
  • Give the Windows main thread the 8 MB stack Linux and macOS have by @kleinbem in #2681
  • Make --draft and --draft-max reach the speculative config by @Virgile-pct in #2952

console

  • Drop stale WASM CDN pin breaking image describe by @iamthesvn in #2296
  • Show context length and quant for remote models on client nodes by @iamthesvn in #2262
  • Stable long replies, sticky model choice and free/paid with prices by @michaelneale in #2315
  • Show mesh graph node hover card in fullscreen mode by @iamthesvn in #2355
  • Correct VRAM, quant, and fit values in model drawer by @iamthesvn in #2289
  • Attach images whose clipboard MIME type is empty by @iamthesvn in #2443

config

  • Write the config and the service env file owner-only by @Virgile-pct in #2220
  • Move native runtime settings into Runtime section by @ndizazzo in #2221
  • Keep MESH_LLM_LIFECYCLE_LOG_PARSER out of the saved config by @kleinbem in #2482
  • Warn in config validate on a model_path serve would refuse by @Virgile-pct in #2554
  • Keep a replaced file's ACL on Linux and macOS by @kleinbem in #2863

inference

routing

sdk

  • Resume SSE frame scanning so long OpenAI streams stay linear by @i386 in #2269
  • Allow embedded nodes to stop during startup and downloads by @i386 in #2440
  • Stop an embedded node during startup without waiting out its probes by @kleinbem in #2864
  • Allow immediate restart after stopping embedded nodes by @i386 in #2878
  • Honor OCR output format choices through Accept by @ndizazzo in #2854

windows

Other

Security

Other changes

  • Make the llama upgrade canary platform independent by @i386 in #2811
  • Revert "ci(release): run every release leg on GitHub-provided runners (#2980)" by @ndizazzo in #2984

Internal

CI, build, test, and repository work with no user-facing behavior change (182 changes)

CI and release engineering

  • Restore native libraries for crates publication by @i386 in #2224
  • Catch llama canary failures before family fan-out by @i386 in #2223
  • Register Skippy composition scripts in ownership catalog by @i386 in #2263
  • Keep canary upstream tests green with 16 workers by @i386 in #2270
  • Retain repair agent build and test logs by @i386 in #2271
  • Prepare embedding SDK in canary build job by @i386 in #2292
  • Preserve automation compatibility and Windows shutdown checks by @ndizazzo in #2312
  • Add a Windows CUDA 13 runtime for Blackwell GPUs by @Virgile-pct in #2291
  • Install just before macOS runtime, product, and Swift console verification by @ndizazzo in #2327
  • Fall back to Rust runner cleanup after smoke tests by @ndizazzo in #2333
  • Route workflows and scripts through rust automation by @ndizazzo in #2375
  • Remove python scripts replaced by rust automation by @ndizazzo in #2376
  • Prepare automation before routing affected crates on hosted planners by @ndizazzo in #2402
  • Build automation once per run and reuse cached hosted binaries by @ndizazzo in #2405
  • Clear the compiler wrapper when bootstrapping a cached hosted xtask by @ndizazzo in #2406
  • Restore lane automation in the macOS Swift SDK smoke by @ndizazzo in #2411
  • Remove remaining python tooling by @ndizazzo in #2377
  • Let macOS canaries start with disk-only caching by @i386 in #2415
  • Avoid rehashing NAS models before canary tests by @i386 in #2416
  • Preserve model manifests during canary admission by @i386 in #2419
  • Retain complete canary failure evidence by @i386 in #2420
  • Isolate macOS compiler cache servers per job by @i386 in #2439
  • Run canary native tests once per build by @i386 in #2432
  • Register the skippy-qualification ownership domain and core rows by @i386 in #2426
  • Prevent macOS compiler cache startup races by @i386 in #2445
  • Require preinstalled tools for canary jobs by @i386 in #2450
  • Certify stable llama.cpp releases by default by @i386 in #2451
  • Certify and publish selected Mesh revisions by @i386 in #2462
  • Register the windows-skippy-probes platform row by @Virgile-pct in #2390
  • Run the native Skippy probes on Windows by @Virgile-pct in #2389
  • Publish certified llama upgrades after skipped attempts by @i386 in #2475
  • Report skipped GPU inference rows in the step summary by @Virgile-pct in #2442
  • List the Windows Skippy probe row in the CI inventory by @Virgile-pct in #2467
  • Register the capsules plugin ownership domain by @StevenMih in #2449
  • Register the plugin-capsules domain by @StevenMih in #2474
  • Run the Skippy runtime and serving lib tests in the probe row by @Virgile-pct in #2511
  • Bundle each default plugin's pinned release archive by @ethanyzhang in #2546
  • Build the Windows runtimes for native-abi changes by @Virgile-pct in #2510
  • Run Llama Upgrade on lab Macs by @i386 in #2582
  • Drop the plugin-capsules domain by @ethanyzhang in #2537
  • Install just in the native runtime build job by @michaelneale in #2597
  • Persist verified Windows release ABI builds by @michaelneale in #2662
  • Declare the grammar-sampling recipes in the Just ownership contract by @i386 in #2692
  • Unblock preview cache publication and Swift builds by @michaelneale in #2688
  • Drop ARM-only CUDA architectures from Windows runtimes by @Virgile-pct in #2664
  • Include prereleases in Llama Upgrade by @i386 in #2457
  • Write the Skippy CLI checksum sidecar in the canonical format by @Virgile-pct in #2697
  • Run the posix_path Windows test in the native roster by @Virgile-pct in #2701
  • Route standalone Skippy API changes to Quality by @i386 in #2728
  • Compile a focused controller for planning and validation by @ndizazzo in #2709
  • Validate the standalone Skippy API and Rust bridge by @i386 in #2729
  • Accept verified regular model cache files by @i386 in #2743
  • Support macOS Bash in family certification by @i386 in #2758
  • Move process supervision into its own crate by @ndizazzo in #2711
  • Make rolling preview on demand only by @michaelneale in #2595
  • Reduce Linux test disk usage and refresh caches by @i386 in #2762
  • Keep macOS compiler cache alive through test phases by @ndizazzo in #2799
  • Register the libclang CPU image on the linux-cpu runtime row by @i386 in #2830
  • Admit the libclang-capable CPU runner image by @i386 in #2821
  • Route process-supervision and mesh-llm-routing to platform-windows-cfg by @Virgile-pct in #2778
  • Run process-supervision and mesh-llm-routing in the Windows unit row by @danielwinterw in #2844
  • Sync planner contract fixtures with platform-windows-cfg routing by @i386 in #2845
  • Compile the xtask Clippy graph into the trusted seed by @Virgile-pct in #2849
  • Keep compiler seeds compatible with Clippy profiles by @i386 in #2853
  • Accept automation tools built with an explicit host target by @kleinbem in #2755
  • Allow Llama Upgrade from a repair branch by @i386 in #2816
  • Disable Rust debug information across CI by @i386 in #2870
  • Certify Llama families across available memory pools by @i386 in #2869
  • Restore Quality canary build contract checks by @i386 in #2874
  • Focus llama upgrade repairs and preserve recovery by @i386 in #2879
  • Allow local llama repair outputs under ignored checkout data by @i386 in #2884
  • Admit empty canary runner credentials by @i386 in #2886
  • Make Llama repair progress easier to follow by @i386 in #2875
  • Make generated family check work without ripgrep by @i386 in #2892
  • Return Llama repair agent after focused checks by @i386 in #2899
  • Show native Goose output during Llama repair by @i386 in #2905
  • Limit Linux Llama canary CTest concurrency to four by @i386 in #2909
  • Stabilize validation fixtures on busy hosts by @ndizazzo in #2898
  • Preserve proxy blocking and canary backend contracts by @ndizazzo in #2858
  • Run Llama Upgrade entirely on lab runners by @i386 in #2918
  • Parallelize frozen Llama native builds by @i386 in #2922
  • Verify Llama candidate parity against bundled commit by @i386 in #2926
  • Use frozen controller revision in Llama verification by @i386 in #2933
  • Report untracked Llama verifier paths by @i386 in #2934
  • Isolate early Llama battery scratch output by @i386 in #2935
  • Ignore selected Llama source checkout in controller by @i386 in #2940
  • Initialize notes helper toolchain and cache by @ndizazzo in #2855
  • Keep Llama retries from restarting verification by @i386 in #2955
  • Run the stability harness on the public cpu image by @Virgile-pct in #2757
  • Run the skippy-cli tests in the Windows unit row by @Virgile-pct in #2916
  • Stop canary receipts vouching for racily changed payloads by @danielwinterw in #2961
  • Convert and publish DFlash drafts from a checked-in manifest by @danielwinterw in #2914
  • Run every release leg on GitHub-provided runners by @i386 in #2980
  • Prepare 0.79.0 version metadata by @i386 in #2991

Build and dependencies

Tests

Refactors, docs, and hygiene

  • Prefer updating owning llama.cpp patches by @i386 in #2265
  • Document local choice and score limits by @i386 in #2328
  • Consolidate Mesh thin-wrapper and v0.78.0 parity fixes by @i386 in #2238
  • Make README wordmark readable on light backgrounds by @michaelneale in #2404
  • Run full crate test suites before opening PRs by @iamthesvn in #2447
  • Choose an inference provider explicitly by @michaelneale in #2549
  • Inference audit conformance bar by @Virgile-pct in #2636
  • List the powermetrics plugin by @Virgile-pct in #2718
  • Drop the native Windows workarounds for fixed bugs by @Virgile-pct in #2676
  • Keep tests and builds working in clean distribution build environments by @kleinbem in #2754
  • Generate native FFI bindings with bindgen from the pinned headers by @i386 in #2761
  • Check for sccache on Windows without a module scan by @Virgile-pct in #2813
  • Show installing pay-lightning with a wallet plugin by @michaelneale in #2842
  • Remove the family certification host lock by @i386 in #2812
  • Scope local validation to changed surfaces by @i386 in #2871
  • Decouple generation from frontend wire types by @IvGolovach in #2682
  • Describe a provider's requires, not the removed payment block by @Virgile-pct in #2890
  • Remove Spiral logo from README support section by @matbalez in #2908
  • Clarify supported OCR provider configuration by @ndizazzo in #2861
  • Call out clippy as no test substitute and CI triage order by @i386 in #2877
  • Remove duplicate byte-size formatters by @ndizazzo in #2931
  • Share one definition of mesh stream IDs between clients and nodes by @ndizazzo in #2929

New Contributors

Full Changelog: v0.78.1...v0.79.0