v0.9.0 — static graftd + self-healing install

0.9.0 — 2026-09-02

— 2026-09-02

Added

• heimdall setup — hardware-fitted configuration: detects accelerator
  (Metal/CUDA/CPU) + physical cores, generates annotated ~/.graft/config.yaml
  (existing file backed up, never clobbered), downloads an embedding model
  from a built-in catalog (bge-m3 default; bge-small-en-v1.5,
  snowflake-arctic-embed-s, nomic-embed-text-v1.5) or accepts BYO --model-path,
  and installs/repairs the launchd daemon (com.graft.daemon). Flags:
  --model --model-path --threads --instances --accel --graftd --skip-daemon --detect-only. See docs/setup.md.
• Doctor expansion — heimdall doctor now also validates: config present,
  graftd --check-config passes, embedding model file present + sane size,
  plist valid, daemon running — each with SETUP NEEDED guidance.
• graft-cpp fork + subtree — vendor/graft/ is now a git subtree of the
  graft-cpp fork (pinned v0.1.0-heimdall.2). Fork delta: daemon config
  fallback chain (--config > $GRAFT_CONFIG > ~/.graft/config.yaml >
  defaults, source logged) and graftd --check-config [PATH] (side-effect-free
  resolved-config dump). Update via git subtree pull — see
  vendor/graft/VENDORED.md.
• Self-contained graftd build — llama.cpp fetched via CMake
  FetchContent at pinned tag b10760 and compiled as static libraries.
  graftd has no dynamic libllama/libggml dependency and no
  RPATH/LC_RPATH. Metal shader library embedded on Apple
  (GGML_METAL_EMBED_LIBRARY). CUDA passthrough: -DGGML_CUDA=ON.
• npm postinstall auto-build — npm i -g builds graftd when the
  toolchain is present; prints a SETUP NEEDED block (what failed + log path
  • retry instructions) on any failure and exits 0 — install never breaks.
    HEIMDALL_NO_BUILD=1 skips the build in postinstall and heimdall setup.
• bin/lib/graft-build.mjs — probe (graftd --check-config, exit 0 +
  model_path: in stdout; missing configPath returns error immediately),
  canonical-first find/install (~/.local/bin/graftd probed first; if
  broken/absent and another candidate works — build cache or
  ~/Repos/graft-cpp/build/graftd — it is copied atomically into
  ~/.local/bin/graftd, broken canonical kept as graftd.bak-<timestamp>;
  only if no candidate works does it fall through to build), build (cmake
  configure + incremental build, timeout-bounded), install (atomic copy to
  ~/.local/bin/graftd; graft CLI copied to ~/.local/bin/graft only if
  absent).
• heimdall setup binary handling — finds a working graftd or builds
  and installs one when none exists; --graftd PATH probed before copying
  (broken binary = one-line error, nothing written).
• tests/graftd-binary.test.mjs — asserts the built graftd has no
  dynamic llama/ggml deps, no RPATH/LC_RPATH, and runs after being copied
  elsewhere.

Changed

• bge-m3.gguf relocated from vendor/graft/models/ to ~/.graft/models/
  (models no longer live inside the source tree; setup downloads on demand).
• Setup default instances — always 2 (was 2 only when cores ≥ 8, else 1).
• Linux thread detection — physical cores counted from /proc/cpuinfo
  (unique physical-id × core-id pairs), falling back to lscpu -p=CORE,SOCKET,
  nproc, then os.cpus().
• vendor/graft/ in the npm tarball — vendor/graft/ source, CMake, and
  third_party/{BLAKE3,mpack,sqlite-vec} are shipped; third_party/llama.cpp/,
  build/, models/, *.db are excluded; llama.cpp is fetched at first build
  (network needed once).
• C++ runtime linkage — graftd links stdc++ on Linux, c++ on Apple
  (was c++ unconditionally, which fails to link on Linux/gcc).

Fixed

• dyld: Library not loaded: @rpath/libllama.0.dylib — pre-0.9.0 graftd
  linked llama.cpp dynamically with build-tree RPATHs; the binary broke when
  moved out of the build tree. The static build has no RPATH/LC_RPATH and no
  dynamic llama/ggml dependency.
• Hardcoded third_party/llama.cpp/build dependency — replaced by CMake
  FetchContent; llama.cpp is fetched at configure time from the pinned tag
  and built in place.
• vendor/graft subtree missing 61 source files (src/**,
  include/graft/*.h, VERSION): commit b7845ef's conflict resolution
  deleted them; restored verbatim from the v0.1.0-heimdall.2 squash
  (b4a51de). src/daemon/main.c keeps the heimdall fork delta.
• Tracked dangling symlink removed — vendor/graft/third_party/llama.cpp
  → /tmp/llama-master was tracked in git; removed. The path is now the
  gitignored FetchContent checkout.