Repository navigation
Conversation
The detached update a hook, a reader over budget or the MCP server starts, watch, and serve's poll between questions now run in the background band: PRIO_DARWIN_BG on macOS, nice 19 with the idle I/O class on Linux, background mode on Windows. serve polls on a short-lived thread so its answers keep normal priority; Linux cannot raise a lowered thread back. Foreground commands are unchanged. A refusal is logged and the work goes on. A dense update pays every row an earlier --no-dense write left owed; the sync is pulled out as pay_owed so a test can hold it to that.
The skill moves to skills/zerograph/ with metadata.version set to the release that wrote it, so a re-run by the same release writes nothing and one by another says what it replaced. The repo-query directory an earlier release wrote is removed when it holds only that release's SKILL.md, under any --command; an edited one, a link or one with files of its own is left and named. The skill gains what keeps the index fresh, how to see and pay owed vectors without blocking a turn, and why embed_model belongs in zerograph.toml.
serve joined its lowered poll thread, so a question waiting behind a poll ran at background speed, and the rayon workers were never lowered anyway. Its poll is back on the main thread, and Extractor loses the Send bound the thread needed. On Linux, nice 19 without the idle I/O class is reported as that rather than as no lowering at all.
Four threads on twelve cores instead of three. The balanced row in configuration.md is the four-thread run of 2026-10-09 (taken under the one-third rule, which gave the same count); it has not been re-run on this build.
…tors Generated code, migrations and mocks were a fifth of beauty-crm's embedded rows. skip would also take them from impact, trace, changes and the lexical index; embed_skip leaves those alone, owes them no rows, and the next writer drops the rows they held. Every other row is matched by hash, so changing the key re-embeds nothing else.
onnx-community publishes model_<dtype>.onnx beside model.onnx; the suffix reads that file from the same repository and pinned commit, with the model's own prefixes and pooling. All five embeddinggemma exports load on the CPU provider. On the bench fixture at balanced, none embeds faster than fp32, none loses a paraphrase on the 30-case arm, and q4 is a sixth of the download at 0.91 GB RSS against 1.47 GB. The default stays fp32; the widened 60-case arm could not be re-read.
A variant no longer fetches the fp32 graph's own data file from the profile's `beside` list — `qwen3-embedding:q4` asked for 2 GB of weights that belong to `model.onnx`; the variant's own external data is still found by size. The threads test gets its doc comment back from the embed_skip test inserted above it, and the 0.6.4 notes name `embed_skip` and the `<hub id>:<variant>` spelling.
devmaxxx
changed the title
# Conflicts: # docs/releases/0.6.4.md
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
<hub id>:<variant>A consumer repo ran its git hooks as
update --no-densebecause a dense refresh was feared to load the machine; 18,787 rows ended up owing vectors andasksilently missed recent code. The detachedupdate,watch, and the MCPreindex/switch_modeltools now lower the whole process before any thread starts:PRIO_DARWIN_BGon macOS, nice 19 plus the idle I/O class on Linux,PROCESS_MODE_BACKGROUND_BEGINon Windows. A failure to lower is logged and the work continues.serveand every foreground command keep normal priority, so a question is never answered at background speed. A denseupdatepays rows an earlier--no-densewrite left owed (pay_owed, with a test), so hooks can drop--no-dense.The skill
install-agentwrites is nowzerographand carriesmetadata.version. An oldrepo-queryskill is removed only when it is byte-identical to what 0.6.0–0.6.3 generated; an edited, symlinked or extended copy is kept with a note.Trade-off, recorded in
docs/history.md: the 2026-09-09 removal ofpriority = "background"was about rebuilds paid by whoever waited. This lowers only unattended work, but a foregroundupdatethat finds a background one holding the store still waits behind it;asknever does.balancednow takes 40% of the logical cores (4 threads on 12, the 2026-10-09 four-thread row: build 521 s against 674 s at three).embed_skipglobs keep generated or bulk files in the graph, the lexical index,impactandtrace, but out of the vectors; adding a glob embeds nothing else.<hub id>:<variant>reads a published quantised export; on this CPU none embeds faster than fp32 (docs/bench/2026-10-17-gemma-quantized-results.md), so the default stays fp32 and:q4is offered for memory-constrained machines.Test plan
cargo test— 1608 passed, 0 failed, 8 ignoredcargo clippy --all-targets -- -D warningscleanpriority.rstype-checks for Linux and Windows targets