v1.9.0

Added

• The libraries compiled into the binary ship their notices (#577). git2 builds with vendored-libgit2 and libz-sys with static, so every binary distributed here statically contains libgit2 and zlib, and no package dependency carries their terms. libgit2 is GPLv2 with a linking exception, and that exception is what makes a permissively licensed binary possible at all: it is conditional on the notice travelling with the distribution. zlib’s terms say its notice may not be removed. Neither was shipping, and neither had anything to do with the project’s own license, which is why this is separate from #573 rather than part of it: both were unmet under MIT-only too.

  Both notices now live under third-party/ and reach the release archives, the .deb, the .rpm, the AUR package and the published crate. They are taken from the crate source cargo actually built, not from the libraries’ upstream repositories, because those diverge: libgit2-sys vendors a pinned tree while libgit2 main moves.

  third-party/README.md records which crate version each came from, and that record is checked against Cargo.lock. A dependency bump moves the vendored library without touching a file anyone reads, so without the check the notice would quietly go on describing a version the binary no longer contains.

Changed

• The doc captures show what the binary actually prints (#575). #567 rewrote 165 strings gwm prints and the captures predate it, so several showed text the binary no longer says. They are images, so nothing in CI could notice. 21 moved, and they were behind on a second count the same regeneration fixes: #568’s tilde compression. palette.png shows it best. Its PATH column read five near-identical /Users/kbrdn1/gwm-demo/... rows clipped mid-path, and now starts at ~/gwm-demo/, so what survives the clip is the part that differs per row.

  Five captures came back byte-identical, which is the correct answer rather than a miss: cli-list and cli-agents are CLI output, where paths stay absolute by design, and none of the five carries a rewritten string.

  Two tapes generate.sh does not run are now named in the script and the capture README, because neither absence was reported and a run therefore ended with a tick over a set that still held them stale. demo.tape is the documented one and was run separately here. github-linking.tape is the other, and it is the one capture in the set that nobody but the maintainer can reproduce: it needs a repo with an open PR, which the demo fixture has not, so it points at a hardcoded checkout and photographs whatever it is doing. Both were re-run by hand for this cut, github-linking.png from the release branch itself, which is the only state in which that pane has anything to show. Making it deterministic stays a design question rather than a regeneration.

  The version chip is what set the order of this release. It comes from CARGO_PKG_VERSION, so all 24 TUI tapes bake the version into the image, and a set regenerated before the bump would have advertised 1.8.0 for the whole 1.9.0 line. The cut therefore bumps first and regenerates after, on one branch, rather than following the usual order of merging the work and adding a release commit on top. Every capture published here reads gwm 1.9.0.

• gwm is now dual-licensed under MIT OR Apache-2.0 (#573). Nothing is taken away: MIT stays, in full, under its own file. What is added is the option of Apache-2.0, whose §3 is an express patent grant from every contributor. MIT has no patent clause at all, and that gap is the one thing a corporate legal review reliably stops on. It is also the ecosystem’s default (rustc, cargo, clap, serde, ratatui and git2 all ship this exact pair), so packagers recognise the layout without asking.

  LICENSE.md becomes LICENSE-MIT, copyright line untouched, and LICENSE-APACHE joins it carrying the upstream Apache text verbatim, appendix placeholders left unfilled. Users pick either one, and never have to say which.

  Both texts now travel in every artefact: the release tarballs and zips, the .deb and .rpm payloads, the AUR package, and the published crate. The declaration itself had to be written six times in six syntaxes, none of them interchangeable: MIT OR Apache-2.0 for crates.io and Arch, any_of: for Homebrew, a | separator for Scoop, [ licenses.asl20 licenses.mit ] for nixpkgs. A walk over packaging/, Cargo.toml and flake.nix now fails the suite if any channel declares one half without the other, and a second guard ties the files the release workflow stages into the tarball to the files the AUR PKGBUILD installs back out of it.

  The contribution terms are stated in CONTRIBUTING.md and at the foot of the README: a contribution is dual-licensed the same way unless its author says otherwise. There is no CLA and nothing to sign.

  A license change is worth seeing in the version, which is why this release is a minor and not a patch.

• A message gwm prints reads without an em dash (#567). The rule this project writes under is that published prose does not use the em dash, and the binary’s own output is as published as the README. #516 swept docs/ and #543 finished the skills; neither reached src/, so a user hitting an exec error read a dash the documentation for that same feature no longer used. 165 of them across 161 string literals, in every error message, status line and TUI hint the binary carries, and 49 more in gwm --help, which clap builds out of the doc comments on the CLI types. The completion scripts carried the same text and are fixed with it: gwm completions zsh went from 23 to none.

  The connector was chosen at each call site rather than substituted. A colon where the dash introduced the remedy, a semicolon where the clause already carried a colon, a full stop where a second colon would have made the sentence unreadable. In the TUI the dash was often not punctuation at all but a separator, and there is already one in use there, so / filter, the create form hints and the loader detail now read the way the segments beside them already did.

  Comments are untouched, roughly 1900 of them: a doc comment quoting a spec or a command’s real output has to stay verbatim, and it is read in the source rather than printed. Two tests hold the rule from here on. One scans src/ with a Rust literal scanner rather than a grep, since telling a literal from a doc comment is the whole difference between a guard that holds and one nobody can keep green. The other walks gwm --help and every subcommand’s, because that is where a doc comment stops being a comment, and no scanner can answer that question as reliably as reading what the binary prints.

  Several doc captures showed the old wording and were tracked separately; #575 regenerated them in this same cut, so the published site never serves the two halves out of step.

• The one-line description reads without an em dash in every package listing (#567). #567 swept src/, and its guard is scoped there, so the same tagline kept its dash in eight published fields no test could see: the crates.io description, the deb extended-description, the rpm summary, both flake.nix descriptions, and the Homebrew, Scoop and AUR blurbs. A ninth sat in the flake’s shellHook, printed to whoever runs nix develop. A package listing is read by a user, which is the surface #567 is about. The connector is a colon: the dash was introducing an expansion of the noun before it. Comments in those same files keep theirs, which stays #567’s to sweep.

• The Settings panel sizes to its active tab (#569). #550 gave every bounded overlay one width policy; height stayed a flat percentage of the frame, so the panel took 60% of the terminal whether the tab under it had 3 rows or 173. On a 40-row terminal that is a 24-row box for the Worktree tab’s three fields, roughly six rows of it blank.

  The box is now the header, the body, the footer hint, the border and the padding, clamped between a floor of 11 rows (the shortest tab that carries a real form) and a ceiling that leaves about 25 rows of body. It therefore changes size as tabs are cycled, which is the deliberate trade: the tabs are genuinely different lengths, and the alternative was blank rows on three tabs out of five.

  The ? overlay and the command palette keep their percentage. Measured, they carry about 220 rows and 52 commands, so sizing to content would resolve to the ceiling in every ordinary state and only ever engage on a heavily filtered palette, where it becomes a live resize while typing.

  The exact-height modals (create, rename, both delete dialogues) keep sizing the way they did, border flush with the frame and all. They have no scroll path, so the policy’s two rows of margin would not shrink those boxes, it would take rows off the bottom of them: a delete confirmation for a target carrying a branch would lose its Delete Branch row on a 16-row terminal.

• One spelling for a worktree path, everywhere it is printed in full (#568). The header rendered $HOME as ~ and the table printed the same value raw, so one path appeared twice on one screen in two spellings. The column paying for it is the one that can least afford it: PATH is Fill(1), it takes whatever the other columns leave and by design vanishes first, and it spent 13 of its columns on the home directory on every row. Measured on the demo fixture at 103 columns the column gets about 22 cells, and all five rows read /Users/kbrdn1/gwm-demo, identical, hard-clipped mid-path with no ellipsis. They now start at ~/gwm-demo/, so what survives the clip is the part that differs per row.

  Compression runs before the terminal sanitiser, not after, and that order is load-bearing: the prefix is matched byte for byte against dirs::home_dir(), so sanitising first would rewrite whatever $HOME itself carries and compression would silently stop firing for exactly the users whose home is hostile. The tail is still sanitised, so a worktree directory name cannot ride the tilde into the cell.

  The sidebar’s Path row had the other half of the problem and now shares the same helper: it compressed but never sanitised. Nothing leaked, because ratatui drops a zero-width formatting character rather than painting it, but the row silently showed a path the filesystem does not have.

  gwm path --format=json is unaffected and stays absolute: this is a rendering change in the TUI only.

Docs

• The frontmatter says what the published page should say (#579). The frontmatter of docs/ is the source of the <title> and the <meta name="description"> on gwm.kbrdn.dev: the bridge carries it through verbatim, so a defect there is only fixable here, and it ships green because nothing in the build reads a description.

  Three of them are closed. 3.cli/index.md and 3.cli/1.reference.md said the same sentence to a word, in both locales, so two separately crawled URLs competed for one snippet; the section landing now describes the section it routes to rather than the reference it links. Both roadmap descriptions ran to 297 and 307 characters by enumerating every version line since v1.0.0, which put what is shipping now behind the truncation; they stop at the current line. Getting Started was the only title-cased page in a sentence-case tree and disagreed with the sidebar label and the OG card table the site builds, both of which say Getting started; the title, the navigation.title, the H1 and the ten cross-references that spell the page name follow. The French .gwm.toml page also started on a lowercase schéma, which the bridge was quietly capitalising on publish.

  Two of the three are now pinned by tests/docs_frontmatter_tests.rs: a description over 250 characters, and two pages of the same locale describing themselves the same way. Title casing is left to the author, since separating Getting Started from GitHub issue / PR linking needs a hand-written list of proper nouns that goes stale on the first page named after something not on it.

Fixed

• Tilde compression fires on Windows, and with a trailing separator on $HOME (#568). The home prefix was matched byte for byte, which failed in two ways that had been silent since the helper was written, so the header and the sidebar rendered absolute paths for the affected users rather than ~.

  On Windows the two sources spell the same path differently: a worktree path comes from libgit2, which emits / there, while the home directory comes back with \, so C:/Users/alice/repo never matched C:\Users\alice. Separator spellings are now compared as equivalent, on Windows only, since a backslash is an ordinary character in a Unix directory name.

  HOME=/home/alice/ is legal and is handed back with the separator intact, which left the match ending mid-boundary and refused. Trailing separators are now trimmed before the comparison.