Skip to content
glep

glep / Project

Changelog

glep release history from v0.1.0 to v2.0.0, where each version was published, and what changes when you upgrade from 0.3.1 to 2.0.0.

On this page
  1. First principles
  2. Where versions are published
  3. 2.0.0 (2026-10-26)
  4. Performance
  5. Matching and output
  6. Files and the index
  7. Agent hooks and tooling
  8. 0.3.1 (2026-09-30)
  9. 0.3.0 (2026-07-23)
  10. 0.2.3 (2026-07-19)
  11. 0.2.2 (2026-07-17)
  12. 0.2.1 (2026-07-17)
  13. 0.2.0 (2026-07-17)
  14. 0.1.2 (2026-07-16)
  15. 0.1.1 (2026-07-15)
  16. 0.1.0 (2026-07-15)

This page follows the repository’s CHANGELOG.md. Dates are the tagged commit dates. GitHub releases start at v0.3.1; earlier versions have tags only.

First principles

glep’s version number changes only when a release is cut. 2.0.0 follows 0.3.1 directly: the changes made after 0.3.1 were prepared as 0.4.0 and released as 2.0.0, so there is no 0.4.0 release. Run glep --version to see which one you have.

Where versions are published

  • PyPI: every version from 0.1.0 to 0.3.1, then 2.0.0.
  • crates.io: 0.1.1 to 0.2.3, then 2.0.0. Versions 0.1.0, 0.3.0 and 0.3.1 are not on crates.io.
  • GitHub releases: v0.3.1, with the wheels attached, and v2.0.0, with the wheels and standalone binaries attached.

2.0.0 (2026-10-26)

ripgrep parity and a parallel search pipeline: most of ripgrep’s flags with ripgrep’s semantics, checked byte for byte against ripgrep on both the single-thread and the worker-pool search paths, and the first crates.io release since 0.2.3. This is a condensed summary; the full entries, with measurements, are in CHANGELOG.md.

Upgrading from 0.3.1 (see also Installation):

  • Existing indexes rebuild once, on the first query (postings format v2 to v5).
  • Run in a subdirectory of an indexed tree, glep uses the nearest .glep/ above it and scopes the search to the current directory. 0.3.1 built a second index in the subdirectory.
  • -l, -c and --json are last-wins output modes, as in ripgrep. 0.3.1 rejected -l or -c with --json, and -c with -l, as a usage error.
  • Binary detection follows ripgrep’s reader (a NUL in the first 64 KiB, where 0.3.1 looked at 8 KB), and UTF-16 files with a byte-order mark are transcoded and searched instead of skipped.
  • .rgignore files apply, and --no-ignore-vcs also turns off .git/info/exclude and the global gitignore, so results can change in trees that rely on them.

Performance

  • Worker-pool search: one reusable searcher per worker thread and an in-order printer thread, so output stays byte-identical to rg --sort path at parallel speed. The freshness diff runs sharded in parallel.
  • On macOS the search pool defaults to the performance-core count. GLEP_THREADS=N overrides the default, and -j N still wins. Elsewhere the default stays one worker per core.
  • Search workers read each file of up to 1 MiB whole into one reused buffer instead of memory-mapping it. Bigger files are still mapped.
  • The first query on an unindexed tree holding more than 64 MiB no longer waits for the build. It answers with a full scan, then builds the index in a detached low-priority process. GLEP_SYNC_BUILD=1 keeps the inline build and GLEP_BG_BUILD_MIN_BYTES moves the threshold. Windows keeps the inline build.
  • --files scoped to path operands, or run from a subdirectory, lists from a live sweep of its operands and never opens, builds or writes the index.
  • On macOS, root-wide --files listings are answered from a cache keyed on directory modification times (mtimes), with the glob, type and depth filters applied inside the listing.
  • A search no trigram can narrow chooses its plan before the index opens. Over the whole tree on macOS it searches the listing cache’s file set; on a tree with no index it walks live and builds nothing.
  • Candidate selection orders trigrams by posting-list length without decoding them and decodes only the rarest list. Narrowed searches got about 5% faster end to end.
  • Dotfiles the sweep flags hidden are no longer re-checked against every ignore file unless something can still show them.

Matching and output

  • New flags with ripgrep semantics, among them -w, -x, -S, -s, -v, -m, -M, -o, -N, --heading, -q, -P and --engine, -E, -r, --passthru, --count-matches, --files-without-match, --include-zero, --stats, --stop-on-nonmatch, -a, --binary, --null-data, --crlf, --pre and --pre-glob, -z, --column, -b, --vimgrep, --trim, -0, --path-separator, -H, -I, the separator flags, -p, --color, --colors, --hyperlink-format, --sort and --sortr, -j, --generate, --type-list and --pcre2-version. See the CLI reference.
  • The Rust regex and PCRE2 (Perl Compatible Regular Expressions) matchers are configured like ripgrep’s. --engine auto falls back to PCRE2 for each pattern the default engine rejects.
  • -e is repeatable and -f reads patterns from a file (-f - from standard input). All patterns are ORed.
  • A repeated flag is last-wins instead of a usage error, and ripgrep’s negated and alias spellings are accepted. A flag that takes a value takes the next word even when it starts with a dash, so the agent hooks’ -e '-> Result' works.
  • The output modes -l, --files-without-match, -c, --count-matches and --json are last-wins. -l and -c no longer conflict with --json.
  • Binary files behave as under ripgrep’s default reader, with a 64 KiB first buffer. UTF-16 files with a byte-order mark are transcoded and searched. glep status reports skipped (binary/oversized/utf16).
  • Errors follow ripgrep: an unreadable file or directory, a failing --pre command, or a missing operand makes the exit code 2 while the search goes on. A reader that closes the pipe early ends the run quietly instead of printing Broken pipe and exiting 2.
  • --count-matches, -c -o and the --stats match count find matches the way ripgrep’s summary printer does.
  • The printers are grep-printer 0.3, ripgrep 15’s, so -r combines with --json.

Files and the index

  • Ancestor index discovery: glep finds the nearest .glep/ upward, or uses GLEP_INDEX_PATH, scopes the search to the current directory, and prints paths relative to it.
  • Path operands scope the freshness sweep. Only full sweeps advance the --ttl clock or trigger compaction, and a scoped query within the window skips its sweep.
  • ! whitelist rules in ignore files rescue dotfiles, .rgignore files are honored, and existing indexes rebuild once (postings format v2 to v5).
  • -g honors ! negations with last-match-wins ordering. New: --iglob, --glob-case-insensitive, -T, --type-add, --type-clear, --ignore-file and -d. Filters follow ripgrep’s walker level by level.
  • Path operands behave as in ripgrep: a named file is searched unconditionally, several operands group in command-line order, and ignored or symlinked directory operands are walked live.
  • New live-scan options that never touch the index: --no-index, -L, --one-file-system, --require-git outside a repository, the --no-ignore-* toggles, and the -u, -uu and -uuu ladder.
  • --no-ignore-vcs also turns off .git/info/exclude and the global gitignore, as in ripgrep.
  • The planner is a recursive required-literal extractor that fuses adjacent literals and stays linear in the pattern. Case-insensitive narrowing follows Unicode case folding.
  • Parallel first queries on an unindexed tree all answer, with a live scan, instead of failing on the lock. A query that runs while another glep holds the lock answers for the tree its own sweep found.

Agent hooks and tooling

  • The Claude Code and Cursor hooks map Grep’s multiline to -U, turn an offset of K and a head_limit of N into | tail -n +K+1 and | head -N pipes, and leave content searches with -n: false to the built-in tool. The Cursor hook also reads case_sensitive: false as -i and include as the glob.
  • bench/bench.sh compares against ripgrep at its default parallelism with matching output and ignore flags, and the new bench/compare.sh is the kernel-tree benchmark and differential harness.
  • The command-line interface (CLI) runs on a thread with an 8 MiB stack, so Windows debug builds no longer overflow it.
  • PCRE2 is built from source and linked statically (.cargo/config.toml sets PCRE2_SYS_STATIC=1), so the release wheels need no system PCRE2 library.

0.3.1 (2026-09-30)

  • Package metadata now points to https://velofy.co/glep/ and https://github.com/velofy/glep (homepage, documentation, repository and changelog links on PyPI; 0.3.1 was not published to crates.io).
  • The README was rewritten, and the old GitHub Pages site now forwards to this documentation.
  • No code changes since 0.3.0.
  • Adds the CHANGELOG file.

0.3.0 (2026-07-23)

Escape hatches.

  • New --hidden flag: include dot-prefixed files and directories. The index now stores hidden files with a flag, so the flag needs no rebuild. .git is always excluded.
  • New --no-ignore flag: search gitignored and .ignored files with a live scan that never opens, updates, or writes the index.
  • --json now ends with ripgrep’s summary event.
  • -l now conflicts with --json.
  • glep status counts only binary and oversized files as skipped.
  • With --no-ignore, index and status are treated as patterns.
  • Cursor integration: cursor/install.sh and a hook that emits both Cursor and Claude Code output formats.

0.2.3 (2026-07-19)

  • Path filters given as ./path or as absolute paths are normalized to index-relative paths.
  • README and site: full CLI documentation, package links, social cards, updated numbers.

0.2.2 (2026-07-17)

  • Windows support, with a Windows continuous integration (CI) job and a Windows wheel on PyPI.
  • PyPI classifiers.

0.2.1 (2026-07-17)

  • macOS: the freshness sweep uses getattrlistbulk, with a fallback to the portable walker. GLEP_NO_BULK_SWEEP forces the walker.
  • The sweep collects results per thread instead of taking a lock per file.
  • New GLEP_TIMING environment variable for per-stage timings on stderr.
  • The bulk sweep honors all gitignore sources; a .ignore file triggers the walker fallback.
  • The release workflow can be started manually.

0.2.0 (2026-07-17)

  • New -c / --count mode, with ripgrep parity. The Claude Code hook maps Grep count queries to -c.
  • New -A and -B context flags.
  • New -U / --multiline mode.
  • glep status updates the index before reporting.
  • Per-file read errors during a search print a warning to stderr instead of failing.
  • New wheels: macOS x86_64, Linux aarch64, Linux musl x86_64.

0.1.2 (2026-07-16)

  • Performance: the matcher is compiled once, -l stops at the first match in each file, and output streams in chunks.
  • Queries where nothing changed no longer write to the index.
  • First published kernel-scale benchmark numbers.

0.1.1 (2026-07-15)

  • README links and images use absolute URLs so they render on PyPI.
  • The crates.io package excludes the site, docs, and assets.

0.1.0 (2026-07-15)

First release.

  • Trigram index in .glep/ with a memory-mapped postings file, a manifest, incremental updates through a delta segment, and full-rebuild compaction.
  • Self-healing freshness sweep on every query, --ttl, and --max-filesize.
  • Content search with ripgrep-format output, -i, -F, -l, -g, -t, -C, --json, and -e.
  • --files glob listing with gitignore-style glob semantics.
  • glep index and glep status.
  • Claude Code skill and PreToolUse hook.
  • Wheels on PyPI for macOS arm64 and Linux x86_64.