Skip to content

v7.30.0

Choose a tag to compare

@github-actions github-actions released this 04 Sep 19:21
· 2194 commits to main since this release
dce4eb1

Engine identity unchanged: engine-service-v0.1.100.

Fixed

  • Local mode no longer embeds client-side when the engine embeds
    (nexus-b7s8t). nx index repo and nx index rdr ran the bge ONNX model
    on every chunk while the service-backed vector client discarded the
    vectors and the engine embedded again. On one engine, in a 600 s window
    on the same repo, the 7.29.1 client landed 38 files at 894% CPU and
    8.2 GB resident; this client lands 336 files at 100% CPU and 305 MB.
  • nx commands no longer load onnxruntime they never use (nexus-ct17r).
    nexus.pdf_extractor imported MinerU at module scope, which pulled in
    magika and onnxruntime on every nx store put; ORT's telemetry then
    raced interpreter teardown ("recursive_mutex lock failed", Abort trap 6).
    MinerU is resolved on the first PDF extraction, and the fastembed probe
    no longer imports it.
  • The av exclusion reaches user installs (nexus-heykz). uv reads
    [tool.uv] override-dependencies from the invoking project, never from
    the wheel, so every user install carried PyAV and its ffmpeg-62 dylibs
    beside opencv's ffmpeg-61. The overrides file now ships in the wheel and
    nx self install hands it to uv; the fresh-install probe asserts av is
    absent. A bare uv tool install conexus still carries av until
    nx self install moves the box onto the generation layout (README).
  • Git bundles and data containers are never indexed as prose
    (nexus-b9m7a). A .bundle whose first 13 KB is a text ref list passed
    the 8 KB binary sniff, registered a catalog document on every run,
    produced no chunk and was re-stamped by every HEAD change, which doctor
    reported as an unfenced producer. The sniff reads 64 KB, and .bundle,
    .pack, .idx, .npz, .npy, .parquet, .pkl, .pickle, .onnx
    and .safetensors are skipped by extension.
  • MinerU per-page workers no longer leave orphans (nexus-5ny9r). The
    worker's multiprocessing pool child and resource tracker survived every
    page reparented to init (324 on one box after a 38-page run) because the
    post-exit sweep resolved the group from an already-reaped pid. The sweep
    now uses the session id recorded at spawn. A cached "no live MinerU
    server" verdict expires after 60 s with one health probe, so a server
    started mid-run is picked up instead of every remaining page paying the
    subprocess cost; a version-skew verdict never rechecks.
  • nx census reviews reports the current repo's post-commit hook state
    (armed, stale, not installed, unmanaged) first, using the same comparison
    nx doctor makes and honouring core.hooksPath, instead of asking whether
    the hook is installed; a truncated review record states characters
    reviewed of characters total. Review records are now listed by the
    reviewer's agent attribution (server-side filter) before the first-line
    check, retiring the whole-project download. RDR-201 amended after the intrastate
    reanalysis: the lifecycle table header no longer claims an omitted status
    goes uncaught, the checker's guard-independence limit is documented, GH
    #1402 is pinned to its refusing cell, and the inversion incident's failure
    class is asserted while the incident itself is stated as outside the
    table's inputs.

Changed

  • Dependencies (nexus-jpsn1): python-pptx, opencv-python-headless and
    uuid7-standard are dropped (nothing imported them); docling is
    declared as docling-slim[convert-core,format-pdf,models-local] at the
    same locked 2.125.0 (the meta-package's standard extra added rapidocr,
    faker and polyfactory that nexus never used); uvicorn loses its
    [standard] extra (nx-mcp is stdio-only, and watchfiles cost 115 ms on
    every MCP spawn). 22 distributions leave the lock. A dev venv that
    carried python-pptx beside MinerU's pypptx-with-oxml (both own pptx/)
    needs one uv sync --reinstall-package pypptx-with-oxml.
  • Release gates time their index runs (nexus-98zsp). The sandbox shakedown
    and the shakeout compare seconds per chunk against a committed
    per-corpus, per-box-class baseline at a 2x ceiling and print the
    engine's ONNX thread and admission lines beside the number; a missing
    baseline is recorded and reported, never a pass. v0.1.99's 8x embed
    slowdown had passed every gate because none read a clock.
  • The test substrate provisions its PG bundle once per box under xdist
    (nexus-9rnfr), serialised with a lock beside the per-tag cache, and
    retries a sigstore TUF-refresh failure under it.