v0.77.0

[0.77.0] - 2026-09-28

Added

skippy

• Expand and streamline llama family certification by @i386 in #1865
• Recover llama family certification through Inkling by @i386 in #1880
• Add run-ahead verify windows by @danielwinterw in #1409
• Cut over to graph-derived stage filtering by @i386 in #1927
• Bound exact-state recorder admission with two-class credits by @i386 in #1936
• Add draft fallback for N-gram misses by @danielwinterw in #1410
• Serve non-chat GGUF workloads via OpenAI API by @IvGolovach in #1833
• Surface graph reuse counters to the host by @danielwinterw in #2016
• Replay extracted stage programs by @i386 in #2025
• Wire next-generation KV cache and publisher defaults by @i386 in #1838
• Serve Laya decision models on /systemone by @danielwinterw in #2083

openai

• Publish Terminal events on the routing node by @StevenMih in #1668
• Serving provenance, usage and request_digest on the host-served terminal event by @StevenMih in #1841
• X-mesh-target / x-mesh-exclude remote-mesh routing headers (+ x-mesh-served-by echo) by @StevenMih in #1671
• Thread weights_digest onto serving_provenance by @StevenMih in #1942
• Response, tool_calls and reasoning digests on the terminal event by @StevenMih in #1946

payments

• Payments engine as in-process payments.v1 plugin (alternative to #1926) by @michaelneale in #2035
• Bound payments.v1 bookkeeping ops and pin the provider per exchange by @i386 in #2047
• Project advertised pricing and recovery through payments.v1 by @i386 in #2048
• Block providers that take the input payment and deliver nothing by @michaelneale in #2057

Other

• Weights_digest — SHA-256 of served GGUF bytes, alongside identity_hash by @StevenMih in #1708
• Carry an optional self-reported claimed log head on PeerAnnouncement by @StevenMih in #1709
• Add a live runtime event stream with a contract-pinned wire format by @ndizazzo in #1777
• Explain the announced capacity in gpus by @Virgile-pct in #1805
• Add package generation defaults and reasoning budgets by @i386 in #1878
• Genesis invitations for private mesh, and correcting naming of models for remote access by @michaelneale in #1896
• Adapt Anthropic Messages through shared host ingress by @i386 in #1911
• Prototype OpenJEV System One on Skippy by @michaelneale in #1939
• Read back the peer's X-Capsule-Id on RemoteMesh terminal events, labeled peer-asserted by @StevenMih in #1944
• Let a private-mesh service carry the invite token by @i386 in #1960
• Configure Hermes and OpenClaw without launching by @michaelneale in #1964
• Add --parallel to set serve lanes from the command line by @danielwinterw in #1973
• Add anonymous opt-out usage reporting by @danielwinterw in #1993
• Plan the local fit and auto-join on GPU memory, with gpu.host_ram_offload as opt-in by @Virgile-pct in #2019
• Finish the structured native runtime event pipeline by @ndizazzo in #2039
• Advertise System One capability by @i386 in #2093

Changed

• Replace family flags with typed activation frontiers by @i386 in #1887
• --auto-balance speed-balanced layer placement with closed-loop rebalancing by @danielwinterw in #1935
• Borrow the remembered stage slice instead of copying it by @danielwinterw in #2007
• Drop unsupported Intel Apple targets by @i386 in #2096

Fixed

skippy

• Restore certified llama canary validation by @i386 in #1853
• Certify split architectures by @i386 in #1893
• Downgrade qwen4exp cache hits to full-state snapshots by @i386 in #1912
• Prevent activation overreads and accept loader tensor reshapes by @i386 in #1922
• Allow decoder and MTP profiles to share resident weights by @i386 in #1933
• Emit native MTP generation in v2 package writer by @ndizazzo in #1937
• Batch activation exports by request offset, never by lane order by @danielwinterw in #1971
• Memoise stage slice selection with exact dependency keys by @danielwinterw in #1995
• Avoid duplicate model load in state handoff by @i386 in #2032
• Preserve resident KV cache with durable L3 by @i386 in #2067
• Reuse most of a prompt that exceeds the resident prefix cap by @i386 in #2072
• Let DiffusionGemma System One start and read again by @danielwinterw in #2088
• Preserve explicit F32 KV cache type by @sofiageo in #2089
• Reuse backend sampler across requests by @i386 in #2080
• Reuse backend sampler across reset by default by @danielwinterw in #2094

runtime

• Register termination-signal handling once at runtime entry by @i386 in #1969
• Size the shared KV pool by budget, never by the lane override by @danielwinterw in #1974
• Install against the current host Skippy ABI by @i386 in #1987
• Validate speech audio metadata and keep an emptied host set retryable by @i386 in #2006
• Reserve the multimodal projector in the fit plan by @i386 in #2074
• Shut down cleanly on SIGTERM instead of hanging until SIGKILL by @ndizazzo in #2054
• Address #2039 review findings by @ndizazzo in #2078
• Require shutdown signal registration before startup by @ndizazzo in #2077
• Avoid duplicate model log summaries by @i386 in #2085

mesh

• Expose upsert_served_model_descriptor to production callers by @Virgile-pct in #1885
• Keep departed peers from being resurrected by stale gossip by @i386 in #1970
• Set claimed_log_head in the stale-announcement fixture, main is red without it by @Virgile-pct in #1980
• Derive peer serving state and routing eligibility from observed liveness by @ndizazzo in #2058

Other

• Bound layer artifacts and publish replacements atomically by @i386 in #1718
• Skip zero-dimension GGUF tensor placeholders by @i386 in #1807
• Read Windows adapter VRAM from a 64-bit source by @Virgile-pct in #1836
• Preserve plugin models during health grace by @michaelneale in #1837
• Install composed product-v2 bundles in the self-updater by @i386 in #1844
• Clear stale llama.cpp CMake cache on generator mismatch by @ndizazzo in #1846
• Redact the home directory on Windows too by @Virgile-pct in #1852
• Reject rooted POSIX paths in delete on Windows too by @Virgile-pct in #1854
• Preserve launched model names and remove default briefing by @michaelneale in #1879
• Enforce Linux glibc compatibility by @ndizazzo in #1881
• Avoid rehashing cached Hugging Face GGUFs by @i386 in #1892
• Read process facts on Windows instead of answering Ok(None) by @Virgile-pct in #1899
• Launch Claude Code with inherited stdio on macOS by @i386 in #1908
• Return Responses-shaped 502 body from MoA failures (#1905) by @i386 in #1910
• Normalize flat function tools before chat validation by @i386 in #1913
• Use served model context limits in launchers by @i386 in #1914
• Floor the node reserve at the compute graph; explicit starting cut by @danielwinterw in #1938
• Respect canary runner caches and repair Kimi-K3 certification by @i386 in #1958
• Accept layer packages across producer ABI versions by @i386 in #1959
• Inspect local models without writing download caches by @i386 in #1963
• Keep nested runtime settings in their own TOML tables by @i386 in #1975
• Box GET route futures so the router frame stops summing its arms by @Virgile-pct in #1976
• Update upstream pin to 26394b4e67 by @i386 in #1977
• Build the tests on Windows, and close the rooted-path hole they expose by @Virgile-pct in #1978
• Certify upstream 8212c78024 by @i386 in #2033
• Redact the Windows home directory inside log words by @Virgile-pct in #2043
• Clear the Windows-only clippy warnings behind Unix-gated code by @Virgile-pct in #2044
• Report a held config write lock on Windows, and spawn a portable test child by @Virgile-pct in #2050
• Compile the GPU identity guard in test builds by @ndizazzo in #2053
• Keep MoA off peers that charge for the model by @michaelneale in #2060
• Hide Logs and Configuration on client-only nodes by @ndizazzo in #2063
• Stop reporting feature bits 38 and 39 as reserved by @Virgile-pct in #2065
• Clear Windows clippy errors and name a held cache root by @Virgile-pct in #2066
• Stop dropping operational audits behind the fallback recursion guard by @i386 in #2069
• Keep the GitHub star count current by @i386 in #2091
• Preserve free targets on mixed paid meshes by @michaelneale in #2097
• Regenerate model capability bindings by @ndizazzo in #2098

Other changes

• Decide there is nothing to regroup after the link pass by @Virgile-pct in #1871

Internal

CI, build, test, and repository work with no user-facing behavior change (89 changes)

CI and release engineering

• Check out commit-convention script from the default branch by @i386 in #1824
• Resume partial crates.io publishing by @i386 in #1826
• Remove daily layer package queue by @i386 in #1825
• Skip confirmed crates when resuming by @i386 in #1828
• Identify crates.io status probes by @i386 in #1829
• Link crate verification to release runtime by @i386 in #1830
• Add downloaded models to llama canary by @i386 in #1700
• Keep llama canary repair active until gates pass by @i386 in #1827
• Allow parity source variant rows by @i386 in #1832
• Catch SDK runtime JSON breaks before main by @michaelneale in #1695
• Authorize runtime event model for PR and main by @ndizazzo in #1842
• Configure llama canary git identity by @i386 in #1861
• Run llama canary repair with Goose by @i386 in #1862
• Recover release-note entries credited only to a roll-up by @ndizazzo in #1835
• Render llama canary agent output as text by @i386 in #1867
• Simplify agentic-replay workflow to essentials by @i386 in #1717
• Register build accelerator routing by @i386 in #1872
• Complete promoted runner transition by @i386 in #1873
• Retry rate-limited ABI cache saves by @i386 in #1877
• Restore the trusted Linux host sccache seed for amd64 host rows by @i386 in #1889
• Allow large family certifications to finish by @i386 in #1890
• Seed release Linux host objects by @i386 in #1891
• Scope console-print ratchet to product code by @ndizazzo in #1840
• Retire the console-print ratchet for a plain output gate by @ndizazzo in #1860
• Avoid duplicate agent family certification sweep by @i386 in #1894
• Own .gitattributes in the tooling rule by @ndizazzo in #1898
• Flag unused dependencies with cargo-machete by @iamthesvn in #1895
• Add the Windows unit row's steps, inert until the catalog registers it by @Virgile-pct in #1917
• Register the windows-unit platform row by @ndizazzo in #1918
• List the Windows unit row in the platform-checks slice summary by @Virgile-pct in #1920
• Give repaired llama candidates a full verification budget by @i386 in #1919
• Use configured Goose provider for upstream canary by @i386 in #1925
• Run agentic replay nightly with guarded regression repair by @i386 in #1928
• Unblock nightly replay input verification by @i386 in #1930
• Pair dense and recurrent product smokes by @i386 in #1931
• Run nightly replay with its required Python environment by @i386 in #1932
• Run llama canary families as parallel jobs by @i386 in #1949
• Register Mesh and Skippy product layout ownership by @i386 in #1961
• Support canary reruns and submit smaller models first by @i386 in #1979
• Certify trusted MeshLLM revisions with mesh_ref by @i386 in #1982
• Preserve source-owned canary handoff contracts by @i386 in #1984
• Route canary families by available memory by @i386 in #1988
• Clean self-hosted outputs and relax canary memory reserve by @i386 in #2000
• Accept pretty-printed canary evidence by @i386 in #2027
• Add optional PR workflow canary by @ndizazzo in #2004
• Serialize canary host lock contention by @i386 in #2031
• Resolve Rust test and Clippy batches against the revision under test by @i386 in #2024
• Support Mesh and Skippy source layout migration by @i386 in #1983
• Preserve certified canary handoffs across reruns by @i386 in #2034
• Isolate canary runner resources by @i386 in #2049
• Require every cfg-divergent crate to be Windows-routed or listed as unverified by @Virgile-pct in #2040
• Register six split-certification family artifacts by @i386 in #2011
• Isolate canary host resources by @i386 in #2052
• Keep release composers artifact-only by @ndizazzo in #2079
• Simplify llama canary pipeline by @i386 in #2090
• Prepare v0.77.0 source by @i386 in #2100

Build and dependencies

• Remove unused dependencies flagged by cargo-machete by @iamthesvn in #1869
• Make sccache and fast linkers repository defaults by @i386 in #1870
• Check out every text file with LF on every platform by @Virgile-pct in #1883
• Align macOS deployment targets with llama.cpp by @i386 in #1929
• Drop deprecated ggml_mul_mat_set_prec from the patch queue by @i386 in #1989
• Make packaged native runtimes self-contained and pin-isolated by @danielwinterw in #2015
• Remove the transitional host-runtime glob re-export by @iamthesvn in #2030
• Prepare llama on Windows with prepare-llama.sh and key the build dir by pin by @Virgile-pct in #2028
• Resolve a relative MESH_LLM_LLAMA_DIR against the repository root by @Virgile-pct in #2038

Tests

• Render fixture model paths through toml_edit by @Virgile-pct in #1834
• Stop test guards from using the real home on Windows by @Virgile-pct in #1847
• Write test keystores to the temp dir, not the real one by @Virgile-pct in #1850
• Render fixture paths through TOML at six more sites by @Virgile-pct in #1888
• Keep the digest cache out of the real home under test by @Virgile-pct in #1882
• Keep the smart-auto StartNew tests off the network by @Virgile-pct in #1916
• Give the strict absolute-path case a path that is absolute everywhere by @Virgile-pct in #1900
• Make the weights_digest cache-hit test deterministic by @StevenMih in #1947
• Gate the symlink escape test to Unix like its sibling by @Virgile-pct in #2002
• Build short-lived invoices after the slow setup (engine-suite flake) by @i386 in #2046
• Wait for agentic replay runtime metadata by @i386 in #2084
• Fail the feature-bit guard on a FEATURE constant it cannot read by @Virgile-pct in #2086

Refactors, docs, and hygiene

• Remove the legacy stage preparation protocol by @i386 in #1780
• Simplify sharing Ollama, vLLM and LM Studio through plugins by @michaelneale in #1823
• Label --local-model-only as a debugging endpoint; clarify --gguf vs --model by @StevenMih in #1820
• Give announcement building and merging their own module by @Virgile-pct in #1806
• Thank spiral.xyz for project support by @michaelneale in #1863
• Route console output through the sink-aware writer by @ndizazzo in #1848
• Retire runtime console prints for the writer, tracing, and typed events by @ndizazzo in #1851
• Write inline progress directly to the stderr handle by @ndizazzo in #1858
• Follow-ups from #1841 review by @StevenMih in #1962
• Make the clippy gate pass on Windows by @Virgile-pct in #1997
• Distinguish join tokens from private mesh admission by @michaelneale in #1968
• State the real runtime selection boundary by @i386 in #2070

New Contributors

• @iamthesvn made their first contribution in #1869
• @sofiageo made their first contribution in #2089

Full Changelog: v0.76.2...v0.77.0