gwm roadmap

August 15, 2026 · View on GitHub

This document tracks where gwm is heading. It complements CHANGELOG.md (what already shipped) and the open issues (the source of truth for scope details).

Each item below links to its GitHub issue. The scope, alternatives considered, and acceptance criteria live there. This file is the map, not the spec.

Current state: v1.9.0 stable

The current stable line is v1.9.0 (2026-08-16), the licensing line: gwm is dual-licensed MIT OR Apache-2.0. Nothing is taken away, MIT stays in full under its own file; what is added is Apache-2.0's §3, an express patent grant from every contributor, which MIT has no clause for at all and which is the one gap a corporate legal review reliably stops on. It is also the ecosystem's default pair (rustc, cargo, clap, serde, ratatui and git2 all ship it), so packagers recognise the layout without asking. Alongside it, the libraries statically compiled into every binary distributed here now ship their notices: libgit2 is GPLv2 with a linking exception, and that exception, which is what makes a permissively licensed binary possible at all, is conditional on the notice travelling with the distribution. The rest of the line is the text the binary prints: 165 em dashes across 161 string literals, plus 49 more in gwm --help, which clap builds out of the doc comments on the CLI types; the Settings panel sizing to its active tab instead of to a flat percentage of the frame; and one spelling for a worktree path everywhere it is printed in full.

v1.8.0 (2026-08-13) was the density line: the TUI became compact by default, panes and sidebar sections dropping their box rules for a filled one-line header, buying back two rows and two columns each, with the Status block folded onto a single row and [tui] layout = "bordered" as the opt-out. Alongside it, every bounded overlay resolved its width through one policy instead of four, two of which used to get narrower as the terminal widened, and four separate truncators moved from counting characters to counting the terminal cells their callers actually hand them.

v1.7.1 (2026-08-12) was a documentation patch on top of v1.7.0 (2026-08-12), the largest feature line since 1.0: per-worktree notes, the rich PR/Issue view and its inline review comments, container execution on exec profiles, multi-row selection with a batch delete, and the Symfony preset. It closes the four capability gaps that came out of the comparative read of the field, so the comparison page reports parity where it used to report a deficit. v1.6.1 (2026-08-04) closed the gaps left by the v1.6.0 security fix affecting every earlier version (see the highlights table). The machine-readable contracts frozen at 1.0.0 still hold: the CLI subcommands / flags / exit codes, the --format=json schemas, the daemon JSON-RPC protocol, and the .gwm.toml section set will not break without a major bump (see Stability & compatibility). A default layout is explicitly outside that promise: the policy leaves "any visual detail of the ratatui interface" free to change in a minor, which is why compact by default is 1.8.0 and not 2.0.0.

Since the 1.0.0 milestone (2026-06-26): the 1.0.x patches hardened the line (security-only 1.0.3 among them); 1.1.0 shipped the first outside-report features (#363: persisted sidebar layout, OSC 52 yank that works over SSH) and 1.1.1 fixed global config resolution on macOS; 1.2.0 was the distribution push (#383: Scoop, .deb / .rpm, AUR, aqua, with winget wired and pending upstream); 1.3.0 shipped the agent session pane (#408) and its follow-ups, including the Windows named pipe transport for the daemon (#439); 1.4.0 completed the ? help overlay (#453) and closed the TUI-polish trio (#436 / #437 / #438); 1.5.0 made gwm multi-forge (#419), adding a Forge trait and a GitLab (glab) backend behind it; 1.6.0 fixes GHSA-fffq-vg6f-gxqm and lands the naming-flexibility line. See Shipped highlights.

1.0.0 promotes the entire 0.10.0 train: the rc.1 Settings-editability + TUI enrichment cycle; the rc.2 train (embedded PTY overlays, the TUI keymap redesign, rebindable contextual modal keys + the Settings Keys tab, multi-repo workspace mode, config presets for gwm init, the JSON API + daemon, the current-PR CI indicator, and the file-explorer Working Tree pane); the rc.3 train that closes the GitHub loop: gwm review <PR#> and gwm statusline; and the rc.4 fleet-chore pair gwm exec / gwm clean. The post-rc.4 stable delta completes the 1.0 commitment: the frozen, versioned machine contracts (#317), the published stability policy (#318), the frozen exec / clean surface decision (#319), and the additive features it anticipated: named [exec] / [clean] profiles, bounded --jobs, and --workspace fan-out (#324, #326), plus TUI exec / clean overlays (#325). See the Shipped highlights table for the per-issue breakdown and changelogs/1.0.0.md for the consolidated notes. The MSRV is 1.95 (raised by rusqlite's bundled libsqlite3-sys, which declares no floor of its own; MSRV bumps ride a minor per the stability policy). It is held by a CI job that installs the declared floor and both resolves and compiles the committed lockfile against it on all three runners (#491 ✅), because the previous setup let Cargo.toml claim 1.86 for a whole release line while the graph needed more, silently.

The 0.9.x stable line ships:

Native worktree ops via libgit2 (vendored): single binary, no gwq / git CLI dependency.
CLI + ratatui TUI: gwm <subcommand> for scripts, gwm alone opens the interactive interface.
Per-repo .gwm.toml + user-level global config: branch / path conventions, configurable branch types, declarative GitHub labels / milestones, file copies, regex guards (abort or seed-from-example), no-symlink invariants. A user-level ~/.config/gwm/config.toml deep-merges underneath each repo's .gwm.toml (repo wins on conflicts; GWM_NO_GLOBAL_CONFIG=1 forces repo-only). {home} / {repo} / {repo_path} / {repo_parent} placeholders for repo-relative bases.
Config CLI: gwm config get / set / unset / list / validate / path / edit, git-config-style over dotted keys, comment-preserving via toml_edit.
Lifecycle hooks [hooks.*]: declarative pre_create / post_create / pre_bootstrap / post_bootstrap / pre_remove / post_remove phases, per-step on_fail = abort|warn|ignore, --skip-hooks escape hatch, gated by the same when: predicates (file_exists:, cmd_exists:, env_set:, env_eq:, glob_exists:, with !, &&, || composition). Legacy [[bootstrap.command]] is auto-aliased to [[hooks.post_create]].
CLI aliases + Gitmoji convention: [aliases] in .gwm.toml (or ~/.config/gwm/aliases.toml) expand gwm <alias> to argv before clap parses, with gwm aliases list; [gitmoji] mapping powers gwm commit-prefix, gwm types --gitmoji, and an opt-in gwm hooks install commit-msg hook.
GitHub issue / PR templates: [issue_template] + gwm new <type> <desc> (create issue from a form template, then spin up the linked worktree); [pr_template] + gwm pr [--draft] [--base] [--render] with {commits} / {files_changed} placeholders.
Safety daily: --dry-run on gwm remove / gwm prune (preview before destroying); gwm undo + gwm history backed by an operation journal at $XDG_DATA_HOME/gwm/history.toml (100-entry rotation, per-repo filtering) to recover a misfired removal without git reflog.
gwm sync [<pattern>] [--merge]: fetch a worktree's upstream and rebase (or merge) its branch onto it; refuses a dirty tree or missing upstream, aborts a conflicting rebase/merge to keep the worktree usable.
Bootstrap hardening for hostile clones: TOFU trust ledger on .gwm.toml, --allow-bootstrap / --deny-bootstrap, path-traversal rejection, symlink-safe copy/write primitives, load-time regex validation for deny patterns.
Async-task spine (new in 0.9.0): a generic off-thread worker (coalescing + late-result drop) keeps the event loop responsive: the worktree list refresh (f / r), the GitHub issue/PR fetch (F, per-key generation guard fixes a stale-data race), gwm sync (S), and bootstrap (b) all run on it, animating the statusbar spinner instead of blocking. The TOFU gate stays synchronous before any bootstrap spawn.
In-TUI pane-key family 1 / 2 / 3 / 4 (new in 0.9.0): 1 / 2 focus the worktrees / status panes; 3 opens a lazygit-style Command Logs overlay (scrollable transcript of the external commands gwm ran); 4 opens a Configuration panel showing the resolved .gwm.toml with a per-row source column (repo / user / default). . opens the docs in the browser.
Full theme-role coverage (new in 0.9.0): the resolved [theme] is threaded through every draw_* site, with dedicated name / path chrome roles and staged / modified / untracked working-tree roles; all defaults preserved, pinned by tests/tui_theme_audit_tests.rs.
Lazygit-style details sidebar: four bordered subsections (Worktree / Issue · PR / Working Tree / Recent Commits), status-coloured branch names, header status dot tracking linked PR / issue state, 300-commit Recent Commits buffer with the full topology renderer (○ ◎ │ ╮ ╭ ╯ ╰ ┴ ┬ ─).
Measured TUI sidebar perf pass: branch age is cached on WorktreeInfo, filtered_indices is memoised on FilterState, Recent Commits uses a cached libgit2 revwalk keyed by (worktree path, head OID, limit), and commit-graph pipes store git2::Oid instead of heap-allocated hash strings.
Configurable launchers: [git_tui] drives l (default lazygit -p {path}), [review] drives R (presets: lumen / claude / codex / aider / gh, plus free-form command =). Placeholders {base} {head} {path} {diff} with lazy diff materialisation.
GitHub issue / PR linking + auto-detection: branches matching <type>/#<N>-<slug> auto-link to their issue; CLI link / unlink / open / status for explicit overrides; live state surfaces in the TUI sidebar via gh. A branch's PR is also resolved ephemerally (gh pr list --head) and surfaced as detected, never written to git config, so an explicit link always wins (gwm status, the sidebar F refresh, opt-in gwm list --detect-pr).
TUI personalisation: [tui.keys] remappable keymap with multi-key chord support (g g) + gwm tui keys; : command palette overlay sharing the keystroke Action dispatcher; [theme] role-based colours with catppuccin / gruvbox / tokyo-night / claude-dark presets + gwm theme list / show; sidebar stashes mode toggled by s.
[tui.open] dispatch: o key now spawns $SHELL in the worktree by default; opt back to OS file manager via mode = "finder".
y: yank: copy the selected worktree's path to the clipboard (pbcopy / wl-copy / xclip / xsel / clip).
Vim motions: gg / G jump to first / last; Tab swaps focus between the list and the sidebar; j / k / ↑ / ↓ move selection or scroll the focused panel.
Fuzzy filter (/): sticky nucleo-matcher filter on the worktree list; smart-case, AND on spaces, contiguous beats spread-out; same engine powers gwm switch (picker mode), gwm path / cd / remove / bootstrap (fuzzy CLI lookup).
One-line cd: gwm shell-init <shell> wires up a gcd <pattern> (resolve + cd) and bare gcd (picker + cd) for zsh / bash / fish / PowerShell.
Shell completions: gwm completions <shell> for zsh / bash / fish / PowerShell / elvish (static script generated from the live clap argument tree).
Multiplexer integration: gwm tmux <pattern> [-p] and gwm zellij <pattern> [-p] open the worktree in a new window / pane / tab; refuse to spawn outside an active session.
Responsive + polished TUI chrome: the details sidebar stacks under the table on a narrow terminal (< 120 cols) instead of disappearing; V cycles auto → side-by-side → stacked, H / [tui] sidebar_position flips it left/right. Borderless styled header, single-line statusline with reverse-video badge chips, content-sized themed modals (confirm buttons, animated spinner), git-style working-tree colourisation.
gwm doctor: 8 checks (parses / guard refs / when predicates / external binaries / prunable / orphan branches / base writable / unbound quit keymap), exit codes 0/1/2 for CI.
Confirm-overlay countdown: safety countdown on the delete-confirm overlay when p (delete-branch-on-remove) is armed; duration tunable via [tui].confirm_countdown_secs (0..=5, clamped).
State-sliced TUI internals: tui::app::App is decomposed into tui/state/{create_form,filter,confirm,link_prompt,sidebar,github_fetch}.rs, with dedicated tests for each state slice.
Release pipeline: release.yml on vX.Y.Z tags, pre-release.yml on -rc.N / -alpha.N / -beta.N tags, 5-target build matrix (Linux x86_64 + aarch64, macOS Intel + Apple Silicon, Windows x86_64), GitHub Release assets published through the gh CLI with the workflow token (#146 resolved), per-version changelog body sourced from changelogs/<version>.md (hard-fails if missing), pre-release [Unreleased] dupe guard, Homebrew tap update job on stable releases, cargo binstall support, Nix flake at the repo root. CI test matrix runs on ubuntu-latest / macos-latest / windows-latest.
1000+ tests: integration and state-machine tests covering config (repo + global layering), aliases, gitmoji, hooks, config CLI, naming, bootstrap, doctor, GitHub linking + PR auto-detection, launcher, multiplexer, homebrew formula, binstall metadata, pre-commit hook, TUI state slices (keymap / palette / theme / sidebar), undo/history journal, worktree libgit2 integration, release workflow guards, and CLI end-to-end.

See changelogs/0.9.0.md for the full v0.9.0 release notes, and changelogs/ for the per-version archive.

Shipped highlights

For reference (each linked to its closing PR):

Issue	Shipped in	Feature
#18	v0.3.0	Shell completions (zsh / bash / fish / powershell / elvish)
#19	v0.3.0	`gwm cd <pattern>` + `gwm shell-init <shell>` (the `gcd` wrapper)
#20	v0.3.0	`gwm doctor` (initial check set)
#21	v0.3.0	TUI fuzzy filter (`/`)
#22	v0.4.0	`gwm switch` (picker UI printing the chosen path on stdout)
#23	v0.4.0	Tmux / Zellij integration (`gwm tmux` / `gwm zellij`)
#25	v0.4.0	Extended `when:` predicates (`cmd_exists:`, `env_set:`, `env_eq:`, `glob_exists:`, with `!` / `&&` / `\|\|`)
#26	v0.5.0	Homebrew tap (`brew tap kbrdn1/tap && brew install gwm`)
#28	v0.5.0	Nix flake (`nix profile install github:kbrdn1/gwm-cli`)
#30	v0.5.0	TUI confirm-overlay countdown
#47	v0.5.0	`gwm doctor`: skip merged gwm-style branches in the orphan check
#59	v0.5.0	`[doctor].trunks` config knob
#67 (PR #68)	v0.6.0-rc.1	Issue / PR linking: CLI + TUI controls, `gh`-backed live status
#69 (PR #70)	v0.6.0	TUI Details sidebar redesign (four bordered subsections)
#71 (PR #72)	v0.6.0	TUI Recent Commits panel: lazygit-style layout + full topology renderer
#73 (PR #74)	v0.6.0	Lazygit-style sidebar facelift (`Created` row, status colours, `[tui.open]`, `y: yank`)
#75 (PR #76)	v0.6.0	Configurable launchers (`[git_tui]` + `[review]`): keymap reshuffle `r/R → f/F`, new `R`
#77	v0.6.0	Docs restructure into `docs/` tree (Nuxt Content conventions) + README shrunk to landing
#80 / #81 / #82	v0.7.0-rc.1	Configurable branch types, declarative GitHub labels, declarative GitHub milestones
#93 / #94 / #95 / #96	v0.7.0-rc.1	Bootstrap hardening: symlink-safe copies, path traversal rejection, TOFU trust ledger, guard regex load validation
#97 / #98 / #99 / #100 / #101	v0.7.0-rc.2	Static regex lifting, worktree removal ordering fix, stale-branch refusal, argv-injection guards, E2E create/remove/init tests
#102 / #123 / #124 / #125 / #126 / #127 / #128	v0.7.0-rc.2	`tui::app::App` decomposed into focused `tui/state/` sub-structs
#103 / #104	v0.7.0-rc.2	TUI render-loop perf: cached branch age and memoised `filtered_indices`
#105 / #106	v0.7.0-rc.2	Typed error variants and shared constructors/render helpers
#138	v0.7.0-rc.3	`GitHubFetch` cache keyed by issue/PR number; late results dropped after `invalidate()`
#131 / #134	v0.7.0-rc.3	TUI state encapsulation polish for `ConfirmModal` and `FilterState`
#107 / #108	v0.7.0	Measured P3 TUI sidebar perf: cached libgit2 Recent Commits and `Oid` commit graph pipes
#146 / #147 / #112	v0.8.0-rc.1	Release hardening: `gh`-CLI publish + workflow token, pre-release `[Unreleased]` dupe guard, Windows in the test matrix
#86 / #85	v0.8.0-rc.1	CLI aliases (`[aliases]` in `.gwm.toml` + user fallback, pre-clap expansion), gitmoji mapping + `gwm commit-prefix` + opt-in `commit-msg` hook
#31 / #29	v0.8.0-rc.2	Safety daily: `--dry-run` on `gwm remove` / `gwm prune`, `gwm undo` + `gwm history` operation journal at `$XDG_DATA_HOME/gwm/history.toml`
#89 / #88	v0.8.0-rc.3	Config CLI (`gwm config get/set/unset/list/validate/path/edit`, comment-preserving `toml_edit`) + `[hooks.*]` lifecycle hooks (six phases, `on_fail`, `[[bootstrap.command]]` compat)
#83 / #84	v0.8.0-rc.3	GitHub templates: `[issue_template]` + `gwm new`, `[pr_template]` + `gwm pr` with `{commits}` / `{files_changed}` placeholders
#87 / #32 / #33 / #34	v0.8.0-rc.3	TUI personalisation: `[tui.keys]` remappable keymap with chords + `gwm tui keys`, command palette (`:`), `[theme]` role-based presets, sidebar stashes mode (`s`)
#24 / #27	v0.8.0-rc.4	Quick wins: `gwm sync [<pattern>] [--merge]` (fetch + rebase/merge onto upstream, conflict-safe) and `cargo-binstall` support via `[package.metadata.binstall]`
#190 / #188 / #185 / #187 / #180 / #179 / #181 / #175	v0.8.0-rc.5	Personalisation + polish: user-level global config (`~/.config/gwm/config.toml`, deep-merged under `.gwm.toml`), responsive TUI sidebar (`V` orientation / `H` left-right + `[tui] sidebar_position`), `claude-dark` preset + reworked borderless header, modal polish pass (themed content-sized frames, confirm buttons, spinner), single-line statusline, working-tree colourisation, ephemeral PR auto-detection, `{repo_path}` / `{repo_parent}` placeholders
#170 / #210 / #211 / #214 / #169	v0.9.0-rc.1	Theme-role coverage finish: `[theme]` threaded through every `draw_*` site, dedicated `name` / `path` roles for the chrome and `staged` / `modified` / `untracked` roles for the working-tree panel (all defaults preserved, pinned by `tests/tui_theme_audit_tests.rs`); `git2` `0.20` → `0.21` source migration
#217 / #220 / #222 / #224	v0.9.0-rc.2	TUI statusbar / layout / modal polish: contextual statusbar (context chip + animated GitHub-fetch spinner + action log), direct pane-focus keys (`1` / `2`, rebindable `focus_worktrees` / `focus_status`), off-thread GitHub `F` fetch, reworked create-worktree / link / confirm-delete / Issue-PR modals, stacked-by-default sidebar with per-axis split ratios
#231 / #226 / #232 / #255 / #258 / #262	v0.9.0-rc.3	Async-task spine: generic off-thread worker (coalescing + late-result drop), off-thread worktree refresh (`f` / `r`) and GitHub `F` fetch (per-key generation fixes a stale-data race) on the shared spine; in-TUI Command Logs overlay (`3`), Configuration panel (`4`, per-row source attribution), and `sync` action (`S`, off-thread rebase) completing the `1`/`2`/`3`/`4` pane-key family; fuzzy-filter-in-pane-title + command-palette input polish; internal refactor / perf sweep (#235–#244, per-frame sidebar clone dropped)
#256 / #233 / #248	v0.9.0	Stable delta: off-thread bootstrap (`b`) on the spine (trust gate stays synchronous; Report on completion), open the docs in the browser (`.`, rebindable `open_docs`), and a CI-flaky GitHub-detect test fixed at the root (distinct write-once fake-`gh` scripts)
#279 / #257 / #276 / #267 / #283 / #285 / #281 / #287	v0.10.0-rc.1	Settings-editability + TUI enrichment: editable Settings panel (`4`) with category tabs (Theme / Worktree / TUI / All), per-layer selector (`L`), live-persist TOML; reusable `LoaderWidget` (delete modal consumer); async create-worktree and quit-wait on the spine; Working Tree colour-coded nerdfont counts + row recolouring; Status pane `Diff +ins -del` vs base (three-dot); Issue/PR `●/●` pastilles + nerdfont state-chip badges; cached GitHub state (titles + states survive restarts), initial startup refresh, periodic `[tui].auto_refresh_secs` auto-refresh; herdr-style scrollbars on scrollable modals; flat which-key hints
#35 (PR #289) / #290 (PR #292) / #219 (PR #293) / #294 (PR #297) / #300 (PR #301) / #299 (PR #302) / #36 (PR #303) / #37 (PR #305) / #38 (PR #306)	v0.10.0-rc.2	PTY + power-user + integration train: embedded PTY overlays for lazygit (`l` / `L`) and a native `$SHELL` (`o` / `O`) via `portable-pty` + `tui-term` (MSRV → 1.86); TUI keymap redesign (unified list-view bindings: `p`/`P` pull/push, `c` edit-worktree, `e` exit-to-worktree, `y`/`w` yank, `t` mux pane, `h`/`H` macros, plus`[tui.macro1]` / `[tui.macro2]`); rebindable contextual modal keys under `[tui.keys.modal.<context>]`; edit every keymap live from the Settings Keys tab (keystroke capture + validated write-back); Working Tree pane as a nerd-font file-explorer tree (git-coloured rows, bounded scan); current-PR CI indicator in the Status pane; multi-repo workspace mode (`--workspace`, REPO column, `gwm create --repo`); config presets for `gwm init` (`--preset` / `--list-presets` / `--show`); JSON API (`--format=json` on `list` / `doctor` / `path`) + `gwm daemon` (JSON-RPC 2.0 over a unix socket with `subscribe`)
#308 (PR #310)	v0.10.0-rc.3	`gwm review <PR#>`: materialise an existing GitHub PR into an isolated worktree (resolves head via `gh`, fetches `refs/pull/<N>/head` (cross-fork aware, any PR state) into `review/pr-<N>-<author>-<slug>`, links the PR + points the diff base at `origin/<base>`); safe-by-default (no bootstrap / lifecycle hooks unless `--bootstrap`), `--name` overrides the branch
#309 (PR #311)	v0.10.0-rc.3	`gwm statusline`: the first real consumer of `gwm daemon` (#38), a dependency-free client rendering a compact one-line worktree summary (active branch, count, dirty / ahead / behind, linked issue / PR) for tmux / starship / zsh; `--watch` rides the `subscribe` stream, degrades to a blank line + exit `0` when no daemon is reachable; `docs/5.integrations/4.daemon-consumers.md` (EN + FR)
#313 (PR #314)	v0.10.0-rc.4	Fleet chores across worktrees: `gwm exec [<slug>...] -- <cmd>` (run a command in each worktree sequentially, every non-main worktree by default, or listed slugs; everything after `--` forwarded verbatim; `✓ / ✗` rollup, non-zero exit on any failure) and `gwm clean [<slug>...] [--yes]` (report, or with `--yes` reclaim, heavy build artifacts `target/` / `node_modules/` / `dist/` / `build/`; deletes only git-ignored dirs, never follows symlinks, not journaled into `gwm history`)
#317 / #318 / #319 / #324 / #325 / #326 / #334	v1.0.0	The 1.0 commitment: frozen, versioned machine contracts pinned by `tests/contract_tests.rs` (`SCHEMA_VERSION = 1`, daemon `schema_version`, `docs/schema/README.md`) (#317); published stability & compatibility policy (`docs/6.development/3.stability.md`, EN + FR) (#318); frozen `exec` / `clean` surface decision (#319); named `[exec]` / `[clean]` config profiles + bounded `--jobs` parallelism (#324); `--workspace` fan-out for `gwm exec` / `gwm clean` (#326); TUI exec (`x`) / clean (`X`) overlays (#325); help-overlay + overlay polish (#334)
#363	v1.1.0	First outside-report release: persisted sidebar layout (`V` / `H` write back to config) and OSC 52 clipboard yank that works over SSH
#383	v1.2.0	Distribution push: Scoop bucket, native `.deb` / `.rpm` packages, AUR (`gwm-cli-bin`), aqua standard registry; winget wired, pending upstream merge
#408 (PR #435) + follow-ups #439 (PR #444) / #440 (PR #442) / #441 (PR #443) / #445 (PR #446)	v1.3.0	Agent session pane: detect AI-agent sessions (Claude Code, Codex, opencode, Mistral Vibe) per worktree from on-disk artefacts, surfaced as an AGENT column (table + TUI), an `a` detail overlay with pinning, `gwm agents` CLI, additive JSON `agents` field, statusline segment. Follow-ups: Windows named pipe transport for `gwm daemon` / `gwm statusline` (owner-only pipe, server identity verified by owner SID) (#439); `gwm clean` ENOTEMPTY race tolerance (#440); process-level liveness: a dead recorded PID demotes the session immediately on unix (#441); fixed-height attach prompt with a scrollbar (#445)
#437 (PR #452) / #438 (PR #454) / #436 (PR #455) / #453 (PR #456)	v1.4.0	TUI polish + complete help overlay: Working Tree scroll from the Status context (`J` / `K`, rebindable, viewport-clamped) (#437); responsive sidebar heights via a pure layout solver (guaranteed floors, proportional split, Agents pane never clipped, Working Tree scrollbar) (#438); CI checks overlay: one row per `statusCheckRollup` entry with workflow + duration, `Enter` opens details, `/` filter, `f` refresh (#436); `?` help overlay documents every modal context with a per-section completeness guard, which-key re-audit (`exec` / `agents`), and a reserved-typing contract across every input sub-mode (#453)
#419 (PR #458) + #463 (PR #464)	v1.5.0	Multi-forge: a `Forge` trait with two backends, the existing GitHub one (`gh`) and a new GitLab one (`glab`). Worktrees, bootstrap, branch naming and the `branch.<name>.gwm-*` link storage stay forge-neutral; only the network layer knows which forge is in play. Ships the `forge` key in `.gwm.toml`, a `[forge_hosts]` table read from the user's own global config, `gwm trust add`, `$GWM_GLAB`, and a refusal to assume an unrecognised host is GitHub (which would have sent an authenticated call, and a token, to whatever host a cloned repo named). GitLab specifics absorbed at the parse boundary: `iid`, nested subgroup paths, the `/-/` URL infix, date-only milestones, project-vs-group labels, and a pipeline-to-CI-state map where an unknown status never aggregates to green. Fix: the trust ledger keys on the repo again, not on its host (#463)

| GHSA-fffq-vg6f-gxqm + #415 / #416 / #417 / #418 / #479 + #491 | v1.6.0 | Security fix + naming flexibility. A branch name could inject a command into a lifecycle hook: placeholders were expanded into sh -c unescaped, and git permits ;, |, &, $, backticks and redirections in a ref name, so a branch someone else pushed ran arbitrary commands as anyone who had trusted their own repo's hooks, with no trust prompt in the path (the gate covers the repo's hooks, never the branch name entering them). Affects every version up to and including 1.5.0, no backport. Values are shell-escaped on expansion, env values stay raw because they never see a shell, and hooks additionally get GWM_* environment variables that need no quoting. Alongside it: gwm create --name drops the <type> <issue> <desc> requirement (#416), the TUI create and rename forms present the fields the repo's patterns actually ask for in pattern order, in both directions (#418), and the declared MSRV becomes an honest 1.95 held by a CI job that resolves and compiles the locked graph at the floor (#491). Verifying the sequence produced the Fixed section: terminal-escape neutralisation of echoed config (#473), single-pass placeholder expansion (#494), branch rollback on a failed gwm create (#487), Windows path rules on free-form names (#475), and worktree-aware branch reads (#477) |

| #502 / #506 / #507 + #423 / #511 | v1.6.1 | Bidi follow-up to the security release, and the documentation pipeline. The 1.6.0 neutralisation rests on char::is_control, which covers C0, DEL and C1: it does not cover the twelve characters carrying the Bidi_Control property, which are Cf, not Cc, and reorder how a terminal renders the text around them without ever being a control byte. The pre-trust bootstrap summary inherited the gap, the one output whose job is to let someone authorise a shell command out of an unvetted repo. The TUI worktrees table was never on the path the CLI sinks protect at all: measured on ratatui 0.30, every render path drops the zero-width bytes but List and Table keep the Bidi_Control ones, so a fetched ref could read in an order it is not stored in. Both closed, the neutralisation landing in the width-clipping funnel so a column added later inherits it, plus an alias expansion refused rather than neutralised because it becomes argv before clap is reached. Alongside: the published documentation now resyncs and redeploys itself when main moves (#423, https://gwm.kbrdn.dev), and herdr-plugin-gwm gets an integration page in English and French (#511) | | #515 / #420 / #528 / #421 / #484 / #392 + #521 / #531 / #422 | v1.7.0 | The feature line that closes the comparative read of the field. Four capability gaps came out of reading lazyworktree and gwq against this codebase; the last of them close here, so the comparison page (#422) reports parity where it would have reported a deficit. Per-worktree notes (#515): N opens the selected worktree's note in an editable modal rather than suspending the TUI for $EDITOR, stored as plain Markdown under .git/gwm/notes/ in the main checkout, so it survives gwm remove and stays greppable with gwm shut down; keyed on the branch, with a refusal rather than a silent collision when a volume folds two branch names together. Rich PR/Issue view (#420): I renders the description, author, branch pair, diff size, CI rollup, reviews and conversation at no extra request, the fields riding the call gwm already made for the rollup; the comments anchored to a diff hunk followed as a second transport (#528), GraphQL-only, fired by the view rather than by every gwm status. Container execution (#421): a [container] block on an exec profile, where the mount is the substance rather than the wrapper, since a linked worktree's .git is a file holding an absolute host path and mounting the worktree alone yields a container in which git does not answer. Multi-row selection (#484): Space marks, d deletes the batch behind one confirm that snapshots its targets, with gwm remove a b c as the non-interactive half. Symfony preset (#392), built on Symfony's dotenv convention rather than copied from the Laravel one. Alongside: the TUI delete now runs the remove hooks and records the undo journal (#521), and the removal sequence records itself at its point of no return instead of before the destructive call, which had left gwm undo blocked for the whole repo on a refused removal (#531) | | #524 / #523 | v1.7.1 | Documentation patch, binary unchanged. The agent session pane is what the repository description leads on and it had no page of its own; tui/agent-sessions now covers the four backends, the freshness rules and the pinning, in English and French, and five text-only pages get the capture they described in prose. The demo GIF closes its last defect (#523): the agent pane, which needed a fake on-disk agent store rather than a tape edit, since the fixture simulated no session at all. Shipped with docs_assets_tests, because a capture referenced but absent is not a cosmetic dead link: Astro fails the whole build on it, so the docs site stops deploying. That exact failure happened during this cycle, from the other side: the sync computed capture links from cli/reference.md while writing pages into cli/reference/, one level deeper, so every link was short a ../. Latent since the CLI reference split existed, and only triggered once an image finally lived there | | #545 + #549 / #547 / #550 / #548 / #554 + #560 + #562 + #563 / #553 / #544 | v1.8.0 | The density line, and the width arithmetic underneath it. Compact by default (#545): panes and sidebar sections drop their box rules for a filled one-line header, two rows and two columns back each, with [tui] layout = "bordered" as the opt-out and the same pass moving modal titles into the top rule (#549). The idea predates the feedback that triggered it; the work was never the removal but relocating focus, which the border used to carry and the header now does, plus a section_bg theme role that is an indexed colour rather than a translucent white so the mode survives a terminal without truecolor. [tui] status_one_line (#547) is the content half: the Status block folds onto one row, three more rows for the panes, on by default under either layout. One modal width policy (#550): four rules became one, two of which branched on term_width <= 80 and so got narrower as the terminal widened, while four others had no ceiling at all. A crash on returning from a fullscreen surface (#548), and not a cosmetic one: Terminal::clear snapshots the cursor over DSR, and the return path from a PTY overlay is exactly when a terminal is least likely to answer in time. Four truncators moved from characters to terminal cells (#554 / #560 / #562 / #563), on ratatui's own CellWidth per grapheme rather than unicode-width, which disagrees with the renderer on more than CJK: لالالا is 3 columns to one and 6 painted. ASCII rendered identically throughout, which is what kept it standing. unicode-width is now a dev-dependency, kept as the contrast measure that proves a width fixture is one the two disagree on. Alongside: forms scroll to their focused field instead of clipping it away (#553), and the doc captures were regenerated narrower, fitted to their content and on a darker terminal (#544) | | #573 / #577 / #567 / #569 / #568 + #575 / #579 | v1.9.0 | The licensing line, and the text the binary prints. Dual license MIT OR Apache-2.0 (#573): MIT stays in full under LICENSE-MIT, and LICENSE-APACHE joins it for §3, an express patent grant from every contributor that MIT has no clause for at all. Users pick either and never have to say which. The declaration had to be written six times in six syntaxes, none 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), so a walk over packaging/, Cargo.toml and flake.nix now fails the suite if a channel declares one half without the other. The vendored libraries 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 conditional on the notice travelling with the distribution; zlib's terms say its notice may not be removed. Both were unmet under MIT-only too, which is why this is separate from #573. They are taken from the crate source cargo actually built, and third-party/README.md records which version, checked against Cargo.lock. The em dash sweep reaches src/ (#567): 165 across 161 string literals in every error message, status line and TUI hint, plus 49 in gwm --help, which clap builds out of doc comments, and the completion scripts that carry the same text; the connector was chosen per call site rather than substituted, and the one-line description was fixed in the eight published package fields no src/ guard can see. Two tests hold the rule, one scanning literals with a Rust scanner rather than a grep, the other walking what the binary prints. The Settings panel sizes to its active tab (#569), between a floor of 11 rows and a ceiling leaving about 25 rows of body, where it used to take 60% of the frame whether the tab held 3 rows or 173. One spelling for a worktree path (#568): the header rendered $HOME as ~ and the table printed it raw, so one path appeared twice on one screen in two spellings, in the one column that can least afford it; compression runs before the terminal sanitiser, which is load-bearing, and the fix carried two silent failures with it, / versus \ on Windows and a trailing separator on $HOME. Alongside: the doc captures show what the binary actually prints after #567 and #568 (#575), and the docs/ frontmatter says what the published page should say, with two of its three defects pinned by a test (#579) |

If an issue still shows open on GitHub even though its work shipped, it's a tracking issue waiting for a follow-up audit: check the CHANGELOG and the linked PR before reopening scope work on it.

Next up

1.0.0 is cut (2026-06-26): the 0.10.0 train plus the 1.0 commitment work (frozen contracts #317, stability policy #318, the exec / clean surface freeze #319 and its additive follow-ons #324 / #325 / #326) all shipped to the stable line. See the Shipped highlights table above and changelogs/1.0.0.md for the consolidated notes.

With the machine surface frozen under SemVer, every minor since is the home for additive work: new subcommands / flags, new opt-in .gwm.toml sections, additive JSON fields under the same SCHEMA_VERSION. Anything that would break a frozen surface waits for a future major.

The next feature line is queued from a comparative read of the field (chmouel/lazyworktree and d-kuro/gwq) against the actual gwm codebase. Four capability gaps came out of it. Two have shipped: the agent session pane (#408) in v1.3.0 with its follow-ups, and multi-forge support (#419) in v1.5.0, which landed deliberately ahead of the rich PR/Issue view so that view is born multi-forge instead of being rewritten later. See the table above for both.

What is actually queued now, after the v1.9.0 cut, is smaller than a feature line and mostly came out of one source: unsolicited design feedback from ratatui's maintainer on the awesome-ratatui listing (#544). Three of its four axes are closed; the fourth, the docs site reading flat, lives in kbrdn1/kbrdn-docs rather than here. The capture pass that closed #575 left one defect of its own behind: the stills are written at 1x and the published site scales them up, so they blur on a HiDPI screen (#581). Beyond it: richer note editing (#557), forms compressing their spacing before they scroll (#559), polish on the rich PR/Issue view (#551), the remaining relays (#525) and the distribution channels still open under #383. Translating the docs into German, Spanish and Japanese (#522) is roughly 155,000 words with no native reviewer lined up, so it stays a decision rather than a plan.

The lot below shipped in the v1.7.0 cut (2026-08-12); it is kept here with its reasoning rather than folded into a one-line highlight, because the sequence is the point. It is not ordered by size. Each step sits where it is cheapest to land, which usually means ahead of the thing that would otherwise have to be reopened to accommodate it.

1. Naming flexibility (#415 ✅ → #416 ✅ → #417 ✅ → #479 ✅ → #418 ✅), plus #480 ✅ / #481 ✅ / #482 ✅ / #475 ✅

gwm create requires the full <type> <issue> <desc> triple, and there is no way to name a worktree freely. Working through this surfaced a real defect rather than just a missing feature:

worktree.branch_pattern is configurable and editable live from the Settings panel, and BranchSpec::branch_name honours it. But parse_branch matches a hardcoded regex instead. A user who customises the pattern silently loses issue/PR auto-linking, gitmoji, and the branch-convention check, with no warning connecting cause to effect.

Sequenced cheapest-first, each step shippable on its own:

#415 : warn in doctor / config validate when the pattern is customised (turns a silent failure into a stated one)
#416 : free-form naming via gwm create --name, additive and contract-safe
#417 : derive the parser from the pattern and drop BRANCH_RE, which removes the defect at the root
#479 : rename a free-form worktree, either freely again or into the pattern
#418 : token-driven create form with a live branch/path preview

Everything checked above is on dev and unreleased, so none of it has shipped to a stable line yet. The original sequence is complete; everything else listed below is follow-up work that came out of building the first three, not a widening of the plan.

Two of #418's three asks turned out to have shipped ahead of it, which is worth recording because it is what a long sequence does to its own tail: the live branch/path preview arrived with #217's follow-up and #416, and Ctrl-T with #416. What #418 actually contributed is the field set and the focus order, derived from the patterns rather than fixed at the canonical triple. Its scope note had argued for a generic token-driven form so exotic tokens would be supported rather than dropped, but the token vocabulary is closed at three editable names, so that design and the narrow one coincide except on ordering. A drift guard holds the intent instead.

A checkbox here tracks what is merged into dev, which is not the same thing as what GitHub shows as closed. Closes #N in a pull request body only fires when the pull request targets the default branch, so a merge into dev closes nothing on its own: an issue is either closed by hand once its work lands there, or left open until dev reaches main. Both states appear above and neither means the work is missing.

#478 is already closed inside #417's PR: branch_pattern and path_pattern need not carry the same segments, so a worktree created as gwm create fix 42 login under feat/#{issue}-{desc} keeps its fix only in the directory name, and rebuilding the triple from the branch alone silently renamed that component on every edit.

#479 was the other half of free-form naming: the rename form refused a free-form branch outright, so a worktree named for a spike could neither be renamed again nor promoted into the convention once it grew an issue number. It shipped ahead of #418 rather than beside it, because both rewrite the same create form and running them concurrently buys only merge conflicts. That ordering is what unblocked #418.

Three more came out of reviewing #417 and were split out rather than folded into an already large PR. They touched different files, so they ran in parallel and are all on dev:

#480 : {repo} expands to the workspace display name when a branch is written and to the directory basename when one is read, so in a workspace holding two repos that share a basename a {repo} pattern writes what it cannot read back
#481 : renaming a worktree closes its open pull request without warning, since the remote rename is a delete plus a create and GitHub closes a PR whose head branch is renamed, including through its own rename API
#482 : a segment frozen as a literal is lost when the pattern reorders its placeholders, because the recovery bounds each literal by the canonical type, issue, desc rank against markers whose real order differs

One follow-up of #416 was taken ahead of #418 rather than after it, because it had to land before the feature shipped rather than after:

#475 : a free-form name was not validated against Windows path rules, so < > " | and the reserved device names (CON, NUL, COM1…) passed every check and failed only when the directory was created, after pre_create hooks had run and with the branch already left behind. --name merged into dev two days after v1.5.0 was tagged, so it had never reached a stable line: tightening the validator was additive then and would have been a breaking change once it shipped.

Four bugs found while verifying this sequence end to end are fixed alongside it. None is naming work; they are recorded here because they were found here:

#477 : gwm pr, gwm commit-prefix and gwm bootstrap read the main checkout's branch (or, for bootstrap, no branch at all) when run from a worktree, because discover_repo deliberately walks back to the main working directory and they asked that handle where they were
#494 : expand_placeholders substituted {home} / {repo} first and then substituted what those expansions produced, so a repo whose own name contains {type} made {repo}/{desc} write a type the pattern never mentions. Now single pass, which is also an enlargement: BranchParser::compile had been refusing that whole class rather than mirror a formatter that rewrote its own output, and those patterns round-trip. Fixing the root meant reverting part of #418, which had derived the form's fields after the resolution as a workaround
#473 : a .gwm.toml is data from a repo nobody has vetted, and the commands that read it back skip the trust gate on purpose, so echoing a value verbatim handed that file a terminal escape channel out of a read-only command. Neutralised at each sink rather than each producer, which is what covers checks and rows nobody has written yet. The worst site was not in gwm config at all but in the TOFU prompt, whose bootstrap summary could be made to erase the row naming the shell it asks permission to run; and toml's parse error quotes the offending source line, so gwm list inside the repo was enough
#487 : worktree::add creates the branch before the directory, because WorktreeAddOptions::reference takes a reference that already exists, so any late failure left it orphaned. #474 and #475 each closed one input set; a full disk is not an input set, so the ordering is what actually bounds the class. Three conditions now gate the rollback, and each came from a different way of getting it wrong: this call created the branch, it still points where the call put it, and no checkout has it as HEAD. The last one is the review's: deleting the reference rather than the branch spares a branch.<name> config section the command never wrote, but it also drops git_branch_delete's refusal on a branch a linked worktree stands on, and that residue is a worktree bound to nothing, which no check reports

2. Bulk selection (#484 ✅)

Space marks the active row and d acts on the marked set rather than on a single row. The request came in for bulk cleanup, which is the obvious first consumer, but demand is not why it went first.

It changes what "the current row" means, and the two features behind it both open on the current row: opening on it directly is the advantage the rich PR/Issue view claims over snacks.gh, and a note is attached to whichever worktree is selected. Landing a multi-row selection model after them would have meant reopening both. Landing it before means they are written against the final model once.

#484 (PR #520) : Space marks a row, d deletes the marked set behind one confirm that reports a count instead of listing rows, with D arming the branch deletion for the whole batch. gwm remove a b c is the non-interactive half, and it resolves every pattern before touching anything, so a typo removes nothing at all, which is what the gwm list --format json | ... | xargs -n1 gwm remove workaround could not offer

Scope was negotiated with the reporter rather than assumed, and the answer was narrower than the issue title: only d reads the mark set. Every other verb keeps acting on the cursor row, which would be a footgun if it were silent, so the pane footer carries the mark count for as long as the set is non-empty. Marks are keyed by path, since the fuzzy filter reranks indices on every keystroke and a worktree id is unique only inside one repo, which a workspace merges several of. The filter and the manual f clear them; the background auto-refresh only prunes rows that no longer exist, because a sixty second timer eating a selection still being built would make the feature unusable. cycle_sidebar_layout moved off Space onto z, which is a default worth naming twice: a .gwm.toml binding a chord that starts with z is now a prefix conflict and is refused at load time.

Two defects came out of reviewing it, both in the part that was new rather than in the part that was moved:

the confirm overlay recomputed its batch after a partial failure. worktree::remove prunes the admin entry before it deletes the directory (#98, so a partial failure cannot leave a phantom worktree), which means a removal that fails on the filesystem still drops its own row, the refresh then prunes its mark, and the recomputation fell back to the cursor row. A second confirm deleted a worktree that had never been marked. A batch can now only ever narrow
a worktree id is the .git/worktrees/<id> entry name, and git hands it back to whoever recreates a worktree with that basename. The overlay snapshots its targets and fires after a countdown, and the CLI resolves the whole batch before running any hook, so both had a window in which the id pointed elsewhere. Both go through a checked removal now, and the check sits on the handle the prune acts on rather than in a wrapper that resolves the name a second time

#521 came out of the same review and was deliberately left out of the PR: the TUI delete writes no undo journal entry and runs no pre_remove / post_remove hooks, unlike the CLI. That is older than this feature, but a batch makes it bite harder, since ten worktrees can now leave in one keystroke with nothing for gwm undo to pop.

3. Symfony preset (#392 ✅)

A seventh gwm init --preset. Modelled on the Laravel one, but not a copy of it: Symfony's dotenv convention is the reverse of Laravel's, .env is committed and carries the neutral defaults while .env.local is gitignored and carries the secrets. So the preset copies .env.local and .env.test.local rather than .env, and the no-aws-rds guard seeds from the committed .env instead of an .env.example a Symfony project does not have. var/ joins vendor/ in the no-symlink invariants, because it holds the compiled service container and the cached routes.

Placed here because it looked like the only item in the lot that is not sequenced against the others: it touches src/presets.rs and examples/presets/, disjoint from the TUI files the rest of the line lives in, so it can run in parallel with any of them without a merge conflict.

That held for the merge order and not for the blast radius. Every stack preset puts its commands in [hooks.*] rather than [[bootstrap.command]], and two gwm doctor checks walked the bootstrap commands alone, so this one arrived with a clean report about a file the doctor had barely read: a typo in a hook's when predicate came back as "no when: predicates configured", and a hook invoking a binary that is not installed passed as fine right up to the moment gwm create ran it. Fixed in the same PR rather than deferred, since it applied to every hooks-based config and the preset was only the thing that made it visible. LifecycleHooksConfig::all_steps() now enumerates the six phases through an exhaustive destructuring, so the next consumer cannot quietly read half the config.

4. Rich PR/Issue view (#420 ✅)

Read a pull request or issue in full without leaving the TUI: description, individual checks, reviews, conversation, and inline review comments. Modelled on snacks.nvim's snacks.gh, whose key lesson is that inline diff comments are reachable only through GraphQL, not through gh --json.

The first of the two big TUI features, and it inherits both the detail overlay the agent pane (#408) already paid for and the Forge trait (#419), so it is born reading GitLab too. Where gwm can beat the reference: snacks.gh has no notion of a worktree, so the user picks from a flat list; gwm already knows the worktree → branch → PR → issue chain and can open on the current row directly. That last point is why bulk selection (#484) goes ahead of it: "the current row" has to mean its final thing before two features are written against it.

Shipped, all three steps the issue laid out. gwm asks gh for the description, the author, the reviews and the conversation in the request it already made for the rollup, and I renders all of it. The third step, the comments anchored to a diff hunk (#528), landed as what it always was: a second transport, not a wider field list. Those threads are reachable through GraphQL alone, so they ride a separate request fired when the view opens, which keeps gwm status (same PR fetch, every invocation) from paying for a query it never reads. GitLab answers that it cannot reach them rather than answering an empty list, since the two are different facts.

5. Per-worktree notes (#515 ✅)

gwm holds everything about a worktree except the part only its author can write: where they were. It knows the branch, the linked issue, the diff against base and the agent session, but not what had just been figured out, what is blocking, or what to check before opening the PR. With eight worktrees open that context lives outside gwm, so returning to one after two days means rebuilding it from git log and memory.

A note attached to the worktree, editable from the TUI and flagged in the table. Storage is settled: a plain markdown file at .git/gwm/notes/<branch>.md in the main checkout, which survives gwm remove, stays readable from the main checkout and is never committed. No sweeper is needed, because "the branch is gone" is a question git answers in one call, and gwm doctor reports what is left. Plain text, because a note is something one may want to grep or open without gwm running, which is also what rules git config out despite gwm already keeping three per-branch keys there.

#515 (PR #530) : N opens the selected worktree's note in $EDITOR, a binary marker flags the rows that carry one, gwm note show [slug] reads it back and the --format=json rows carry it in an additive note field. A rename moves the file, and gwm doctor reports a note whose branch is gone

It was meant to come after the PR/Issue view, which pays for the overlay machinery it would reuse. It does not reuse it: the note opens in $EDITOR through the handoff o already had, so the dependency dropped and it landed alongside.

Three decisions carry the rest of it.

Presence means non-blank, not "the file exists". vi over an empty buffer writes one byte, so both exists() and a non-zero length would light the marker for a note nobody wrote. One predicate, in one function, and every surface reads through it. The cost is that the scan reads the files instead of listing a directory, which is nothing next to the Repository::open and revwalk list already does per row.

A branch name is not always a filename. git accepts < > " |, a component ending in . and the Windows device names; Windows accepts none of them, and it silently normalises the trailing dot away, which would make two branches share one note. Those names carry no note anywhere, refused out loud rather than written under a name that means a different branch once the repo is cloned. The test that fixed the on-disk layout is the round trip: the scan re-joins path components with / explicitly, never through to_string_lossy(), which yields feat\#515-x on a Windows runner and would never compare equal to the branch it came from.

N confiscates its prefix space, exactly as z did in #484. It was unbound before, so a .gwm.toml binding a chord starting with N is now a prefix conflict refused at load time. Stated in the changelog and both keybinding docs, and pinned by a test rather than left to a bug report.

Deleting a note is gwm doctor's to report and nobody's to do automatically. Not gwm clean, whose stated safety property is that --yes only removes directories git already ignores, and not gwm remove, since outliving the worktree until the work lands is the whole reason the note sits in the main checkout's git dir.

6. Container execution (#421 ✅)

A container: block on the existing exec / aliases surface, wrapping the command in docker run. Anything exposing a Docker-compatible socket works for free, so there is no runtime to integrate. The cost is not the wrapper, it is the mount: a linked worktree's .git is a file holding an absolute host path, so mounting the worktree alone produces a container in which git does not answer. The reference implementation has exactly that bug, in the middle of its own agent-per-worktree use case. gwm mirrors the host paths and mounts the main checkout's gitdir alongside, which turns this from a column to match into one to win.

It was deferred for want of observed demand. It moves into the line because the comparison page now sits behind it rather than ahead: container execution is a column in that table, and writing the table first would mean publishing a column gwm loses by default.

Built narrower than the reference's nine fields: image, runtime, extra_args. The dropped ones are not missing, they are covered: -w, -e and -v are docker run flags, and extra_args lands after gwm's own so a repeat overrides them. Two decisions the issue asked to be resolved rather than discovered: there is no interactive knob, because gwm exec is a fan-out over N worktrees and a TTY per container means nothing there, while the TUI exec overlay owns a real pty and therefore runs its container with -i -t; and the command is the container's CMD, so an image's ENTRYPOINT receives it as arguments. Refused on Windows, where mirroring a host path into a Linux container cannot work and the .git file would name a drive letter anyway. The block rides a profile only: the inline gwm exec -- <cmd> is the frozen surface (#319) and a config file must not change what an unchanged command line does. [aliases] needed nothing, contrary to what the issue title suggests. A gwm alias is argv substitution towards a subcommand, not a custom command, so t = "exec --profile ci" inherits the container through the profile it expands to.

7. Comparison page (#422 ✅)

gwm against gwq and lazyworktree. Content rather than product work, but it is what decides whether any of the above is seen. gwq is the incumbent by star count and has been dormant since May.

Last of the product steps, and the position is the point: a comparison is worth writing once the features it compares are the ones shipping. That now includes container execution (#421), which is why it moved behind that rather than behind the notes.

Shipped at /comparison, English and French. The issue's premise had to be thrown out first, which is exactly what the warning above anticipated: it listed agent sessions, container execution, GitLab and per-worktree metadata as lazyworktree's lead, and by the time the page was written gwm had all four. Re-measured against the two projects as they are (READMEs and go.mod manifests, 2026-08-12) they are parity, not a gap, and the page says so. Two things came out of the re-read that the issue did not have: neither competitor carries a git binding in its manifest, so the libgit2 point is a verified difference rather than a claim, and the star gap has narrowed from 30-vs-270-vs-456 to 126-vs-281-vs-461. What genuinely stays ahead is narrower and named: lazyworktree's rich worktree metadata (colour, icon, tags), its per-worktree .wt hook files, and seven more months of people finding its corners.

8. Documentation in German, Spanish and Japanese (#522)

Five locales instead of two, immediately before the cut and after the comparison page. The audience data does not ask for this: of the 122 stargazers counted on 2026-08-04 (126 on 2026-08-12), 83 declare a location and no non-English bloc reaches 10%. It is a forward bet, and each language carries a different one. German is the largest non-English bloc already present. Spanish is not visible in the current stargazers at all, which is the point, since it buys reach that does not exist yet. Japanese is the only one with a competitive argument: gwq, the incumbent this project is positioned against, is Japanese-authored, and that community sustains a dense native technical-content culture that generates real inbound.

The site side is tracked separately in kbrdn1/kbrdn-docs#54 and is deliberately decoupled. The sync pipeline carries the locale as a boolean (isFr), and turning that into a locale key can land before a single translation exists, because a declared locale with no content falls back to English. The SEO wiring needs nothing: hreflang alternates plus an x-default on the unprefixed root are already emitted per page.

⚠️ The content is 51,612 English words across 41 pages, so roughly 155,000 at three languages, and no native reader is lined up for any of the three. Worth naming before it is scheduled against a date.

Post-1.7: TUI density, from the ratatui maintainer's feedback (#544)

The traffic that took the repository from 30 to 122 stars came through the ratatui circle, and its maintainer sent back a design read: captures too wide, too much empty space, borders everywhere, weak contrast on the site. #544 tracks acting on it. The work below is on dev, not in a published version: it lands in the next minor.

Shipped there so far:

#545 ✅ + #549 ✅: compact is the default layout. A section is delimited by a filled one-line header instead of four rules, which buys back two rows and two columns per section; modal titles moved into the top rule, two more rows per overlay. [tui] layout = "bordered" restores the pre-1.8 boxes, deliberately untouched by the compact refinements so it stays a faithful restore. [tui] dim_unfocused arrived with it.
#550 ✅: one width policy for every modal. Four different rules, two of which made a modal narrower as the terminal widened past 80 columns, collapse into a single modal_width(term_width, pct, min, max). The floor also makes 80 columns, the width the docs advertise, a size these surfaces were actually sized for: the confirm modal was 49 columns wide there, with its hint row cut mid-word.
#547 ✅: the Status block folds onto one line, under [tui] status_one_line (default on). Four labelled rows for four values of a handful of characters each was the largest waste left in the sidebar once #545 cut the chrome. A knob rather than a compact-mode behaviour, so bordered folds too.
#548 ✅: the cursor-position read that failed on return from a fullscreen overlay.

Still open, all found by the same pass and all cheap:

#553: the two form modals drop a field below 18 rows. The horizontal policy landed with #550; the vertical one did not, so a short terminal silently hides an input rather than scrolling it.
#554: ellipsize_middle counts characters where the terminal counts cells, so a CJK or emoji title overflows the box it was measured into.
#551: polish pass on the rich PR / Issue view (#420), queued behind the density line rather than reopening it.
#544 itself stays open for what is not TUI code: capture width and the site's contrast.

⚠️ The doc captures still show the pre-fold sidebar. Every screenshot under docs/**/_assets/ is generated from a committed tape by docs/_capture/generate.sh, which drives the installed gwm, so the set regenerates in one pass on the maintainer's machine and has to be redone before the cut, not per-PR.

Deferred

#414: process-level agent liveness on macOS / Linux: detection reads transcript mtime, which is activity, not liveness. It lags a session that is thinking or waiting on a long tool call, and a crashed one looks recent for a while. The targeted Unix refinement shipped in v1.3.0 (#441): a Claude Code session whose recorded PID is gone drops to idle immediately. What remains is the general case, deferred by the issue's own title.

Maintenance

#500 ✅: exec_tests flaked on Linux with ETXTBSY: the test writes an executable and runs it, which races against the fork-to-exec window of any other test in the harness that spawns a process. Fixed in v1.6.1 by retrying on that one errno, with the reason written next to it so it does not read as flake-hiding later
The transitive graph needs a manual cargo update from time to time. .github/dependabot.yml sets no allow: dependency-type: all, so dependabot only ever PRs the direct dependencies. Last full refresh: #499, which also took the vendored libgit2 from 1.9.3 to 1.9.6. Do it after an MSRV change lands, never before: a lockfile refresh is exactly how a floor moves without anyone deciding to, so it has to be checked against a floor that is already settled

Visibility

Not feature work, but tracked here because it gates whether any of the above is seen. The launch itself happens at the v1.7.0 delivery, so these are its prerequisites rather than a parallel track.

Measured on 2026-08-04, and it reframes the section: the repository went from 30 to 122 stars in fourteen days, and over that window the referrers put X first by a factor of three, from Terminal Trove and from the ratatui maintainer. Neither was solicited, and nothing in this repository was feeding them. What those relays republish is visual, which is why the first two items below are images.

#523: re-record the demo GIF. Frame 0 is an empty terminal, and on GitHub, on X and in a Terminal Trove listing the first frame is the thumbnail. It also renders gwm 1.0.1 in the version badge, six minor versions behind, and it predates the agent session pane, which is now the line the repository description leads on
#524: visual coverage of the docs: 26 of the 41 English pages carry no image, including the whole CLI section and configuration/gwm-toml.md, the longest page in the tree and the one that answers the long-tail queries the site exists to capture. The agent session pane has neither a page nor a capture
#525: feed the relays, before the launch rather than after, since they are what amplified the project last time: awesome-ratatui first because that is the circle that produced the traffic, then Terminal Trove and the ratatui showcase, then the generic awesome lists. Bing Webmaster Tools is the remaining indexing gap, and it also feeds DuckDuckGo; Google Search Console is done
The comparison page (#422 ✅) sits in the numbered line above as its last product step rather than here, and the translations (#522) follow it
#516 ✅: the em dash is retired across docs/: 1586 occurrences in 78 of the 79 pages, English and French, replaced by whatever connector the dash was standing in for (a colon where it introduced a list or an explanation, a full stop where it joined two independent clauses, a comma or parentheses around an aside). Fenced code blocks are untouched, so 60 survive there. 45 headings changed shape and none was the target of an internal link, checked by resolving the tree's 194 internal anchors against the heading slugs of both the before and after trees. Merged in #518
#423 ✅: the documentation is published at https://gwm.kbrdn.dev and keeps itself there. A merge into main touching docs/, changelogs/ or Cargo.toml posts a repository_dispatch to kbrdn1/kbrdn-docs, which reruns the conversion, commits whatever drifted and redeploys. The in-repo tree stays the source of truth: the site is generated from it, never edited on the other side

Ambitious

Larger investments with strategic payoff. Gated by user demand or a concrete first consumer. The previous "Ambitious" items all shipped in the v0.10.0 rc trains above: config presets (#37) and the JSON-RPC API + daemon (#38) in rc.2, the daemon's "concrete first consumer" as gwm statusline (#309) in rc.3, and the fan-out / disk-hygiene candidates from the post-rc.2 gap review as gwm exec and gwm clean (#313) in the 1.0 line.

Both ambitious bets from the field review have now shipped. The agent session pane (#408), a new detection subsystem with four backends consumed by the TUI, gwm agents, the daemon and gwm statusline, shipped in v1.3.0. The Forge trait (#419), a structural change to the GitHub layer, shipped in v1.5.0 with the GitLab backend it exists to enable as its first consumer.

No other large bets are queued. They land once an issue with a concrete first consumer is filed.

How to contribute

Pick an item that interests you and read its issue for scope details.
Comment on the issue if you intend to work on it (avoids parallel duplication).
gwm create <type> <issue> <slug> to spin up an isolated worktree (the issue auto-links itself, see docs/integrations/github-linking.md).
Open a PR targeting dev following the conventions in CONTRIBUTING.md (Gitmoji + Conventional Commits, tests required, never squash; full docs version at docs/development/contributing.md).
The issue is the source of truth: this roadmap is updated to reflect what ships in each release.

Items marked good first issue (when applicable) are intentionally scoped so a newcomer can land them without a deep dive into the codebase.

Out of scope (for now)

A few directions the project deliberately steers clear of:

Replacing lazygit / gitui in scope: gwm is a worktree manager. Git history surgery stays with the dedicated tools that already do it well; gwm integrates with them rather than competing.
GUI front-end: the terminal is the target. A GUI app would split focus and dilute the design.
Worktree synchronisation across machines: too much surface (state, conflict, networking) for a tool whose value is local responsiveness.

That can change if a concrete use case shows up. Open a feature-request issue with the rationale.