Skip to content

v1.40.0

Choose a tag to compare

@VictorQuilgars VictorQuilgars released this 02 Nov 11:14

Security

  • Built on Electron 45 (Chromium and Node security fixes), from 44. Nothing the app uses was removed: the preload only reaches contextBridge, ipcRenderer and webFrame, none of the Node shims Electron 45 takes out of sandboxed preloads; the Theme Builder window is sandboxed like its opener, so window.open still hands it back. The synchronous safeStorage calls that seal the credentials are deprecated in 45 and go in 46 — they still work, and moving to the asynchronous ones is the next upgrade's work.

Added

  • GitHub Copilot is an AI provider: your Copilot subscription writes the commit messages, explanations and descriptions. It has no key to paste, so its row in Settings › AI Assistant is a switch: turning it on asks the Copilot CLI — run in server mode through GitHub's own SDK (@github/copilot-sdk) — whether it is signed in, and lists the models your account may use, which then join the pickers like any other provider's. The CLI is yours: install it (npm install -g @github/copilot or brew install copilot-cli), run copilot once and /login; it is found on your login shell's PATH, as git is, or named with COPILOT_CLI_PATH. The SDK can bundle a runtime, and it is 171 MB per platform, so the installers leave it out. The CLI is an agent and the app wants a sentence: every request runs with no tools at all, the system prompt replaced, a scratch directory as its working tree — so the repository's instructions are not read and nothing can reach a file — and the session is deleted afterwards, never left in your copilot --resume list. The runtime starts on first use, stays warm for ten minutes, and goes with the app. A missing CLI or a CLI that is not signed in is said on the row, and a request fails once, saying how to fix it, rather than three times. The model picker wears what a message on each Copilot model costs in premium requests — 0× for what the plan includes, 10× in the warning ink — so a ten-fold model is not picked for one-line commit messages by accident.
  • The model is offered on the refs themselves: what is unpushed, a new cut of a branch's commits, a changelog since a tag (#293). A local branch ahead of its upstream offers Explain Unpushed Changes in its AI menu — the branch reading, asked of <upstream>..<branch> instead of the trunk; a branch with nothing unpushed, no upstream or a gone one does not offer it. Every local branch offers Recompose Commits: the commit composer is handed everything the branch carries over its base as one diff, proposes a new cut of it, and applying the reviewed plan takes the branch back to where it forked — its work staged, nothing lost — then commits the plan exactly as a split of the working tree is committed. It runs on the checked-out branch only: the composer commits in the working tree, and switching branches behind your back is not something it will do, so any other branch shows the row disabled, saying to check it out first. Tracked uncommitted changes are refused before the model is asked, a branch that moved since the proposal is refused at apply, and the old tip stays in the reflog (and is named if the apply stops halfway). Every tag offers Generate Changelog of <branch> Since This Tag — <tag>..<the branch you are on>; the branch rather than HEAD, because the changelog is kept under what it describes, and not at all on a detached HEAD. The unpushed reading and a tag's changelog are kept under their range, beside the branch's own reading rather than over it, and inserting a tag's changelog is no longer refused as a deleted branch. No new prompt: all three reuse Explain Branch, Generate Changelog and the composer.
  • A pull request can be found by number, whatever its state, and its row says more. The Pull requests view holds the fifty open requests, and its search box only narrowed those — so a request merged last week could not be found at all, and #12 matched every number with a 12 in it. Typing #123 now narrows to that one request, and when the list does not hold it the view offers to fetch it — open, closed or merged — and opens it in the pull request sheet; Enter does the same, and a number that is no request says so. A request from a fork carries a marker on its row, naming the fork. The row's hover adds its checks, its review decision and its size (additions, deletions, files). The fork rides the list's own request; the other three are not on the list endpoint at all, so they are one GraphQL query for the whole list, asked when the first card opens and held for a minute — never one request per row, and nothing added to the minute-by-minute poll, which stays free while nothing changes. (#291)
  • A key per side bar view. 1 to 9 bring a view of the side bar into sight — Overview, AI, Worktrees, Branches, Remotes, Stash, Tags, Pull requests, Issues, in the order the VS Code panel's rail draws them — switching the column to its List or AI tab first when it has to. Bare digits, because the graph already owns the plain letters and a chord belongs to the system; never while typing in a field, nor under a dialog or a menu. The ? sheet lists them. (#277)
  • The pull request and issue sections say why they are empty. A repository on GitHub with no account connected had no sections at all, the same as a repository with nothing on GitHub — two situations with different fixes and one blank. The sections now say No GitHub account is connected and open Settings › GitHub; a read that failed for any other reason is quoted, with Try again. A repository with no GitHub remote still has no sections, as it always had. (#292)
  • Side bar rows say more of what they are. A branch row shows how long ago its tip was committed, a worktree mark when another worktree has it checked out — the switch git would refuse there — and a tooltip that says what it tracks and where it stands, its open pull request, and that worktree. A tag row shows its annotation; a lightweight tag, which has none, shows none rather than its commit's subject. A stash row names the branch it was made on and its age — read from the stash's own commit, so a renamed stash keeps it. Every fact comes out of the call that already builds the list, one process per list: nothing is asked of git per row as it scrolls into view. (#278)
  • The commit search and the pull-request filter can run on a judgement engine: TypeSafe (Jev). A provider does not have to answer in prose to be one. TypeSafe answers with typed verdicts — a probability, one option out of a set — and joins the catalog on a fourth dialect, serving the two features that select rather than write: search in words and describe a filter. The catalog now says which features a provider serves, and one that names none serves them all, so nobody's settings moved. The per-feature picker offers only the providers that can serve the feature and heads its list with what is built for it, with one line saying what it buys and the pair one click away; the global default never offers a restricted provider, since that pair answers every feature and would fall through, silently, on the seven it cannot serve. Search judges each commit against the words by its position: the engine is never handed a hash, so an invented one is not skipped after the fact, it cannot be written at all — and a search reads a thousand commits instead of two hundred cut to fit one prompt, keeping the fifty most probable, in about three seconds. Filter asks one question per qualifier and the app composes the query from the same vocabulary the field validates, so it is valid by construction rather than by inspection; a description with nothing filterable in it composes nothing rather than an empty query that reads as success. A provider that publishes no model list reads 2 models — key not checked rather than wearing the tick of a key that answered.
  • A filter can say when something was merged or closed. merged: (pull requests) and closed: (both sections) join the vocabulary of saved filters — typed by hand, validated, suggested, and understood when a filter is described: merged since 2026-09-01 used to come back as a query without its date, the section having no qualifier to hang it on. Measured against the judgement engine on sixteen described filters, twice, sixteen of sixteen: merged in the last week, issues fermées cette semaine and the rest land on the right date and on no other. closed: steps aside when merged: is there — a merged pull request is a closed one, and asked separately the two both lit up on merged this year.
  • A remote row opens the remote on its forge, takes the default back, and opens onto its branches. Right-click a remote in REMOTES: Open Repository on Remote, Open Branches on Remote and Copy Branches URL — built from that remote's own URL, so upstream's branches page is upstream's and not the default remote's, on GitHub (Enterprise included), GitLab and Bitbucket; a remote that is a path on disk offers none of them rather than a link to nowhere. Unset as Default Remote takes an explicit choice back — until now it could only be overwritten by another one — and the badge moves to what the default falls back to, origin or the first remote; it is offered only on the remote someone chose, since origin winning by default has nothing to unset. And a remote opens — its chevron, or a click on the row — onto its own branches, drawn by the same function as Branches › REMOTE, so each carries the same card, switch and fetch on hover and the same menu. The default remote's reading, setting and unsetting now live once in the shared git core, which the panel runs too. (#289)
  • A tag row does what the commit it tags can. It could check that commit out, copy its name, push, hide and delete — and comparing it, copying its sha or its message, or resetting to it meant finding its row in the graph first. Its menu now has Compare (with HEAD, with the working tree, Select for Compare and Compare with Selected), Copy (name, short and full hash, the commit's message), Reset Current Branch to This Tag with its three modes, and Solo — Show Only This Tag's History, which the branch rows had alone. They are the branch row's own entries, built by the same builder from the same handlers. The step a tag adds is the one that matters: an annotated tag is an object of its own, with its own sha — the one the tag list shows — so the menu resolves the tag to its commit (refs/tags/<name>^{commit}) when it opens, and every entry acts on that, never on the tag object; the full refname also keeps a branch of the same name from answering for it. Copy Commit Message copies the commit's subject, as it does on every other row — the annotation is on the tag's card. A tag that no longer resolves, or that tags a tree, opens its menu without the entries it could not honour. (#288)
  • An agent can propose a split, and you review it in the composer (#88). The MCP server's new propose_split hands its cut of your uncommitted work — commits, each a message and whole files — to the same Split into commits drawer the app's own AI fills, with the working changes open beside it and a line saying the plan is the agent's. It is measured against the working tree as it is when the drawer opens, not as it was when the agent looked: a file committed since is dropped and counted, a file the plan left out is listed as in no commit. Everything stays editable — messages, which commit a file goes in, the order — and nothing is staged or committed until you press Create. A proposal that arrives missing or unreadable says so instead of opening an empty drawer.

Fixed

  • A side bar list that could not be loaded says so, and offers to try again. The worktrees, the remotes, the reflog, the submodules, the contributors, the agents and the AI stack are read by the side bar itself, and a read that failed left its list empty — which says there are none: the Agents list read none running for two releases over a host that was refusing the call. The error now takes the list's place, quoted as it came, with Try again, and the header carries no count rather than a zero. The counts on the headers come from one rule now, the one the VS Code panel's rail counts with. (#277)
  • A tag row's hash is the commit it points at. An annotated tag showed the id of the tag object, a hash no graph row carries; it shows the tagged commit now. The tag and stash lists are one implementation shared with the VS Code panel, which had drifted: the panel sorted tags by name rather than by version. (#278)
  • A split commit the model gave no message no longer takes its files with it. In Split into commits, a proposed commit with files but an empty message was dropped — rightly — but its files were still counted as placed, so they appeared neither in a commit nor under in no commit, and stayed uncommitted without a word. They now come back loose, or go to a later commit that names them.
  • A repository is on GitHub if any of its remotes is — and a declared Enterprise host counts. Which remote the pull-request and issue views ask GitHub about was decided in seven places, each one origin (or the first remote when there is none) and then is that one on GitHub?. So a repository whose origin is on GitLab and whose upstream is on GitHub was reported as having no GitHub remote — and a repository on the Enterprise Server the user had declared in Settings was never recognised at all: the token was configured and the API base was right, and the answer was still not GitHub, because none of the seven passed the declared host along. It is one rule now, in one place, that both products and every call read: origin when it is on GitHub, otherwise the first remote that is, and a declared Enterprise host counts as GitHub. The API base and the token follow the same remote, so a call can no longer be aimed at one host with the credential of another, and the recent-repositories list names Enterprise repositories too. A host that has not been declared is still not GitHub.
  • A provider named in an old settings file, or in the panel's pin, always resolves to a model. The last resort of the resolution — the setting that predates per-feature pairs — was the one level nothing checked, and it is the level the other three fall onto. A provider id the catalog does not know, or one that cannot serve the feature, came back as itself with no model at all; it now lands on a provider that can answer.
  • A search in words goes to its best answer, reaches the ones past the page, and says when it is not the whole answer. The model's answer comes back best first, and that order was thrown away at once: every hit lit alike, nothing selected, and the page not grown for any of them — so a hit further back than the graph had loaded was a row it did not have, and the count in the field said fewer than the model found. The graph now goes to the most probable hit it can show and grows the page for every hit within reach, the way the diff search already did. And what an answer leaves out is said, because each of these looked exactly like a complete one: The 50 most likely of 250 matches are shown, when the cap cut it; Only the latest 1,000 commits were read, when the history is longer than a judgement search reads; 2 of 10 parts of the history did not answer, when some requests failed and the rest came back. An answer that returns after the query or the repository changed is dropped rather than lit on rows it was not asked about.