mirror of
https://github.com/zed-industries/zed.git
synced 2026-11-19 06:59:22 +01:00
4f862ffbc6efa25303fc843547f60f4fe8929c2f
40564 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
|
4f862ffbc6 |
Cleanup unused workflows and script files (#66024)
Removes a lot of dead code we have floating around here that is no longer used and has not been for ages. If we ever need to check something out here again, we can just revert parts of these. Currently, these just trigger security warnings if anything and thus are dead weight. Hence removing here. Drive-by fix of our collab staging deploys as well. Release Notes: - N/A |
||
|
|
e841324d0d |
Remove unused GitHub action files (#66019)
We currently do not use nor maintain those files and thus their dependencies can get outdated fast. Hence, removing here to keep the repo cleaner. Release Notes: - N/A |
||
|
|
27ea42bf01 |
Update dependencies to address security vulnerabilities (#65997)
## Summary Update locked dependencies to incorporate upstream security and correctness fixes. ## Testing - macOS arm64: Zed build passed; application launch confirmed. - Focused dependency-consumer tests and optional QUIC compilation passed. - Formatting and diff whitespace checks passed. - Linux and Windows were not tested locally. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [ ] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). (N/A) - [x] Tests cover the new or changed behavior. Release Notes: - Updated dependencies to address recent software ecosystem vulnerabilities. |
||
|
|
fda552edce |
agent: Fix reading the entire threads database after every thread save (#66015)
Part of #64897. Also fixes the `threads.db` read traffic and lock stalls reported in [this comment on #59107](https://github.com/zed-industries/zed/issues/59107#issuecomment-5781374977). During long agent sessions, Zed kept reading the whole `threads.db` file in the background. The reporter saw about 150 MB/s of reads. After every save, we reload the thread list. That only needs small fields like the title and dates. But some of those fields were added to the table later, so SQLite stores them after each thread's large `data` blob, and has to go through the whole blob to reach them. On a 148 MB test database with 120 threads, one reload read 37,958 pages, which is the entire file. This adds an index that holds just the fields the thread list needs. SQLite now reads the list from the index and skips the blobs, and the same reload reads 7 pages. The index starts with the dates, so SQLite also gets the threads in newest-first order without sorting, but that part is minor. A test checks that SQLite actually uses the index. Moving `data` to the end of the table would fix it too. But `threads.db` has no migration tracking, and that change would copy every thread's blob on first launch. I went with the index since it's safer and doesn't touch any existing rows. Saving a thread 200 times in a row on that database (release build): | | Before | After | |---|---|---| | Time per save | 22.2 ms | 2.4 ms | | CPU time for 200 saves | 4.4 s | 0.47 s | Memory use stays the same. The index takes about 24 KB, and Zed builds it once on first launch, in the background. That took about 0.1 s for 148 MB. Next, I'd like to update only the saved thread in the list, so we don't reload the whole list after every save. That'll be a follow-up PR. Release Notes: - Improved CPU usage and disk reads while the agent is working, especially with a long thread history. --------- Co-authored-by: zed-zippy[bot] <234243425+zed-zippy[bot]@users.noreply.github.com> |
||
|
|
18512de483 |
ui: Fix scrollbar settings not applying to open scrollbars (#66017)
Changing a scrollbar setting didn't update scrollbars that were already open. You had to close and reopen the terminal or panel to see the new setting. This broke in #52747, and now open scrollbars update right away. `ui` still doesn't depend on the `settings` crate, those changes reach each scrollbar through a small marker global. This also fixes a scrollbar briefly flickering when switched to `auto`, and an `auto` scrollbar still hiding itself shortly after switching to `always`. Release Notes: - Fixed scrollbar settings not applying to already open terminals, panels and menus until they were reopened. |
||
|
|
9e84b066b5 |
fs: Preserve child watches when removing parent subscriptions (#65998)
## Summary Removing a recursive watch subscription could silently stop events for surviving child subscriptions that shared its backend watch. This change tracks coverage explicitly, installs replacement watches shallowest-first before removing the parent, and retains broader coverage if replacement fails without scanning unrelated registrations during teardown. Once replacement coverage is established, an unwatch error is logged without treating the removed watch as reusable coverage, and retained roots remain visible in watcher diagnostics. It addresses part of [FR-68](https://linear.app/zed-industries/issue/FR-68) and follows the issue identified in #60663; this implementation targets the current `OsWatcher` architecture and preserves existing coverage on promotion failure, whereas that PR targets the older architecture and removes the parent watch before attempting replacements. ## Validation - `corgi test -p fs` — 56 tests passed on macOS; parent verification returned the same result from Corgi's cache. - Native child-event delivery after parent removal passed six consecutive runs. - New and extended regressions for unwatch failure, retained-root diagnostics, and unrelated-removal retries fail against the previous PR implementation and pass with the revision. - `./script/clippy -p fs`, `cargo fmt -p fs`, and `git diff --check` passed. - A bounded release-build probe using the real watcher owner with a fake backend measured teardown of 40,000 nonrecursive registrations at 8.9 ms on base, 856 ms on the previous PR implementation, and 10.5 ms with the revision. This does not measure native backend costs. - Linux and Windows execution remain unverified. Relevant removal sequences can still cause repeated failed native promotions; unrelated removals no longer retry them. Release Notes: - fs: Fixed surviving child watch subscriptions losing filesystem events when a covering parent subscription is removed. |
||
|
|
cfb894a5dd |
fs: Normalize folded watcher path identity (#65995)
## Summary Case-insensitive filesystem watches registered with a trailing separator could silently miss events because registration keys preserved the separator while event-path ancestors omitted it. Normalize path components before case folding so equivalent spellings share the same watch key without resolving symlinks or parent components. Add regression coverage for event delivery, shared-watch lifetime, and root and parent-component identity in the existing watcher test module. Related to [FR-68](https://linear.app/zed-industries/issue/FR-68) (partial fix; parent-watch removal and backend churn remain separate). ## Validation - `corgi test -p fs 'fs_watcher::tests'` — 19 tests passed on macOS. - `git diff --check` — passed. Release Notes: - fs: Fixed missed filesystem events for case-insensitive watches registered with trailing separators or equivalent path spellings. |
||
|
|
c18e71c782 |
languages: Fix Go test task cwd on Windows (#65722)
Fixes #64841. gopls.run_tests left /C:/... causing os error 267. Now parses with lsp Uri to_file_path + drive-letter normalize. Release Notes: - Fixed Go Run Test failing on Windows with directory invalid error. --------- Co-authored-by: giocld <giocld@users.noreply.github.com> Co-authored-by: Kirill Bulatov <mail4score@gmail.com> |
||
|
|
f1ccf949d5 |
gpui_ios: Reuse the cached main-thread marker (#65975)
Follow-up to https://github.com/zed-industries/zed/pull/65934#issuecomment-6332065288 Release Notes: - [GPUI] N/A |
||
|
|
574df448ea |
Preserve transport errors in Google and OpenAI Responses streams (#65940)
Preserve typed transport errors in Google AI and OpenAI Responses completion streams instead of converting mid-stream read failures into generic, non-retryable model errors. - Classify streaming body-read failures separately from malformed JSON, preserving the provider name and original error. - Carry typed errors through the completion mappers and direct/OpenAI-compatible callers. Keep the OpenAI Responses mapper compatible with integrations that still supply generic errors. - Apply the same classification to OpenCode's Google-protocol stream. - Leave timeout durations and retry policies unchanged. Release Notes: - Fixed streaming connection failures being classified as non-retryable errors for Google AI and OpenAI Responses model requests. |
||
|
|
566f893ee3 |
open_ai: Handle Responses reasoning text events (#65938)
Handle the official Responses API `response.reasoning_text.delta` and `response.reasoning_text.done` events, used by gpt-oss-compatible servers and custom OpenAI-compatible gateways. - Map nonempty reasoning-text deltas to thinking output, alongside the existing `response.reasoning.delta` compatibility path. - Ignore the done event for display so its full text does not duplicate streamed deltas. - Synthesize reasoning-text deltas from reasoning-item content in non-streaming JSON responses. - Preserve the existing structured reasoning-item replay path unchanged. Release Notes: - Fixed missing thinking output from OpenAI-compatible Responses endpoints that emit reasoning-text events. |
||
|
|
50304070c2 |
helix: Return to normal mode after pasting in select mode (#65492)
## Summary In Helix, if you enter Select Mode, select some text, and Shift-r to replace the selection with the currently yanked text, you are returned to Normal Mode. However, Zed's Helix Mode does not exit Select Mode. Your next motion will begin a new selection. To fix this, I added a check in the catch-all branch of match vim.mode in the Paste Vim::action to return to Mode::HelixNormal if we're in Mode::HelixSelect. I renamed the branch to better reflect what it's for and that closed the original pull request #65152 ## Testing I've extended the existing `test_shift_r_paste()` to include pasting over a selection in `Mode::HelixSelect` and asserting that the new mode is `Mode::HelixNormal`. I've run `cargo test -p vim helix::test::test_shift_r_paste` which returned `test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 572 filtered out; finished in 0.95s` ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [ ] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. Release Notes: - Fixed pasting in Helix select mode (`shift-r`, `space p`, `cmd-v`) not returning to normal mode --------- Co-authored-by: Smit Barmase <heysmitbarmase@gmail.com> |
||
|
|
fd2aedc471 |
agent_servers: Update ACP SDK to 3.5.0 (#65982)
Updates `agent-client-protocol` to [v3.5.0](https://github.com/agentclientprotocol/rust-sdk/releases/tag/v3.5.0) and adopts its fixes for v2 integration feedback. - **Typed v1 fallback:** v2 startup uses `send_initialize`. A `LowerVersion` outcome or a v1-only `-32022` error triggers the v1 fallback directly, instead of recovering the agent's response from the debug log. The debug-log lookup is now only used when the agent exits right after answering and the process-exit error supersedes the in-band result. - **Termination-aware disconnects:** connection-loss checks now include `termination()` and `is_connection_terminated`, so transport or handler failures (not only clean EOF) are treated as a lost connection. Disconnect cleanup still uses `on_close` plus driver completion: `finish_connection` already runs once the driver ends, and `on_close` stays the ordered barrier for clean EOF. Release Notes: - N/A --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: zed-zippy[bot] <234243425+zed-zippy[bot]@users.noreply.github.com> |
||
|
|
32558ab9d5 |
multi_buffer: Fix word diff ranges for collapsed hunks (#65829)
## Summary Closes #65678. Collapsed diff hunks can produce invalid word-diff ranges when offsets calculated against deleted base text are applied to the modified buffer. When the base and modified text differ in byte length, these offsets can land inside a multibyte UTF-8 character, violating the boundary invariant expected by the text rendering system. The issue occurs because `self.show_deleted_hunks` indicates whether deleted hunks can be displayed, not whether a particular hunk's deleted text is currently present in the multibuffer. This change checks for the corresponding `DiffTransform::DeletedHunk` in `self.diff_transforms` before including base-side word diffs for a specific hunk. ## Changes - Check the per-hunk expansion state before projecting `base_word_diffs` into multibuffer offsets. - Preserve deleted-side and modified-side word highlighting for expanded hunks. - Avoid emitting base-side word-diff ranges for collapsed hunks, where the deleted text is not present in the multibuffer. - Add regression tests in `crates/multi_buffer/src/multi_buffer_tests.rs` covering: - `test_word_diff_byte_132_collapsed`: verifies that collapsed hunks do not emit invalid UTF-8 ranges. - `test_word_diff_byte_132_expanded`: verifies that expanded hunks retain deleted-side and modified-side highlighting. - `test_word_diff_mixed_expanded_and_collapsed`: verifies that expansion state is handled independently for each hunk. - `test_word_diff_expanded_cjk`: verifies word-diff highlighting with multibyte CJK text. ## Testing - `cargo test -p multi_buffer -j 1 --lib` — 72 passed. - `cargo test -p multi_buffer -j 1 --lib test_word_diff` — 9 passed. - `cargo clippy -p multi_buffer -j 1 --lib` — passed with no warnings. - `cargo check -p multi_buffer -j 1 --tests` — passed. - `cargo fmt --check -p multi_buffer` — passed. - `git diff --check` — passed. ### Regression verification The collapsed-hunk regression test fails against the original implementation because a generated word-diff range ends at byte 132, which is not a valid UTF-8 character boundary. It passes with this change. The expanded and mixed-hunk regression tests also pass. The complete macOS editor-rendering crash has not been independently reproduced by an editor-level test; the regression tests verify the invalid-range condition at the multibuffer layer. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI checklist not applicable; this PR contains no UI changes. - [x] Tests cover the new or changed behavior. Release Notes: - Fixed an issue where collapsed inline diff hunks could produce invalid word-diff highlights when rendering multibyte UTF-8 text. --------- Co-authored-by: Kirill Bulatov <mail4score@gmail.com> |
||
|
|
340c703bb2 |
project_panel: Preserve clipboard contents when duplicating entries (#60894)
# Objective - Duplicate in the project panel currently overwrites the system clipboard, and also replaces any pending Copy or Cut in the panel. - Fixes #60891 ## Solution - Duplicate now copies the selected entries directly through the project's copy path instead of running Copy then Paste. - That avoids writing file paths to the system clipboard and leaves the panel's pending Copy/Cut state alone. - Copy, Cut, and Paste keep their existing clipboard behavior. Filename disambiguation, recursive directory copies, undo, selection updates, auto-open, and the rename editor still go through the shared transfer path. ## Testing - Added `test_duplicate_preserves_clipboards`, which checks that Duplicate leaves the system clipboard unchanged, still copies the selected entry, does not paste external clipboard paths, and still honors a pending Cut afterward. - Ran `cargo test -p project_panel test_duplicate_preserves_clipboards`. - Existing `test_create_duplicate_items` still covers the rename-after-duplicate UI. - Reviewers can reproduce the original bug by copying some text, duplicating a file or folder in the project panel, then pasting: the original text should still be there. Copy and Cut should still write paths to the system clipboard as before. - The clipboard write is platform-agnostic. I did not do a separate Windows/Linux UI pass beyond the unit test. ## Self-Review Checklist: - [x] I've reviewed my own diff for quality, security, and reliability - [x] Unsafe blocks (if any) have justifying comments - [x] The content adheres to Zed's UI standards ([UX/UI](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) and [icon](https://github.com/zed-industries/zed/blob/main/crates/icons/README.md) guidelines) - [x] Tests cover the new/changed behavior - [x] Performance impact has been considered and is acceptable --- Release Notes: - Fixed duplicating files or folders in the project panel overwriting clipboard contents. --------- Co-authored-by: Smit Barmase <heysmitbarmase@gmail.com> |
||
|
|
389c22e1c7 |
gpui: Make Modifiers::secondary web platform aware (#65970)
This introduces a new PlatformInfo abstraction that allows querying a gpui app's platform from runtime facts. This is necessary for correctness when dealing with web applications, as their host platform there is a runtime fact, not a compile time fact. Release Notes: - [GPUI] Fixed platform-dependent modifiers and keybinding parsing and formatting in web applications by querying the host operating system at runtime. |
||
|
|
04470cd9b5 |
project: Don't send a project update when no worktree was removed (#65189)
## Summary > Machine translation of my original text (Russian, below): > > Zed started to freeze when opening some of the threads created by an AI agent. After such a thread was opened, the cursor in the message input started blinking rapidly, everything stopped, and Zed could not be closed; the only way out was to kill the process. If several Zed windows were open, all of them froze. > > This happened in Zed on Windows with projects in WSL2. Not every thread in the same project froze, some opened fine, so it was hard to understand why it was happening. > > The investigation was done with the help of an AI (Claude), and it also prepared this fix. Its conclusions are below. Original text: Zed начал зависать при открытии некоторых тредов, которые создаёт ИИ-агент. После открытия курсор в поле ввода начинал быстро мигать, всё останавливалось, и закрыть Zed было невозможно — приходилось снимать задачу. Причём если было открыто несколько окон Zed, зависали все. Всё это происходило в Zed на Windows с проектами в WSL2. Не все треды в одном и том же проекте зависали, некоторые открывались нормально, из-за чего было непонятно, почему это происходит. Разбор ситуации делался с помощью ИИ (Claude), он же подготовил исправление. Его выводы ниже. > **AI-generated (Claude):** > > **Cause.** In a remote project the client sends `RemoveWorktree` whenever it releases a handle to a worktree (`Project::on_worktree_released`). On the server, `WorktreeStore::remove_worktree` answered every such request with `UpdateProject`, even when the worktree was already gone. When the client handles `UpdateProject`, `WorktreeStore::set_worktrees_from_proto` re-creates a handle for every listed worktree it does not hold. A remote client holds invisible worktrees weakly (`WorktreeStore::remote(false, ..)`), so that handle is released right away and `RemoveWorktree` is sent again. While the request that really removes a worktree is still queued behind redundant ones, every answer lists that worktree again and produces one more request. > > Loading an agent thread triggers this because `AcpThread::resolve_location` opens the file of every tool-call location with `Project::open_local_buffer`, which creates an invisible worktree for a file outside the project. When the buffer cannot be opened (an image: "Binary files are not supported") the worktree is released at once. A thread whose tool calls read a couple of dozen screenshots outside the project releases that many invisible worktrees almost simultaneously. > > **Change.** `remove_worktree` sends the project update only if it removed a worktree. > > **Measurements.** Stock Zed 1.22.0 client (Linux, calls counted with gdb breakpoints), remote server built from v1.22.0 with and without this change, loading one thread that creates 21 invisible worktrees: > > | | without | with | > |---|---|---| > | `UpdateProject` handled by the client | 4,741 | 57 | > | `Worktree::remote` handles created on the client | 4,702 | 281 | > | `SettingsStore::clear_local_settings` calls | 9,980 | 645 | > | UI thread busy after the thread loaded | 112 s | 6 s | > > In the captured part of a traced unpatched run, 2,839 of 2,840 `RemoveWorktree` requests removed nothing, and 2,565 of them were for the same worktree. A second thread (33 tool-call locations outside the project) went from 54 s to 6 s. > > The client still re-creates and drops handles for worktrees listed in older updates, and `handle_update_project` still recomputes settings for every listed worktree. Both are bounded now and left for follow-ups. > > Possibly related: #65008, #62542 (similar symptoms with WSL remotes and agent threads). Not verified that this resolves them. ## Testing > Machine translation of my original text (Russian, below): > > I checked it on my machine: Zed 1.22.0 on Windows 11, project in WSL2 (Ubuntu). I replaced the server in `~/.zed_server` with one built from v1.22.0 with this change. The thread that used to freeze Zed every time now opens normally. Original text: Проверил у себя: Zed 1.22.0 на Windows 11, проект в WSL2 (Ubuntu). Подставил в `~/.zed_server` сервер, собранный из v1.22.0 с этим изменением. Тред, который раньше каждый раз вешал Zed, теперь открывается нормально. > **AI-generated (Claude):** > > - New test `test_removing_missing_worktree_does_not_send_project_update` in `remote_editing_tests.rs` fails without the change (`left: 1, right: 0`) and passes with it. > - `cargo test -p remote_server --lib`: 57 passed. `cargo clippy -p project -p remote_server --lib --tests -- -D warnings`: clean. > - The measurements above were taken on Linux (Ubuntu in WSL2, Xvfb), with the project opened as an SSH remote through a local stand-in for `ssh` (client and server on the same machine). > - In the check on Windows described above, the thread replays 710 messages; Zed's log shows it loaded three times with the patched server running, and the connection to the server was idle afterwards. > - To reproduce: open a remote project, let an external agent read about 20 image files outside the project in one thread, then reopen that thread. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [ ] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). (no UI changes) - [x] Tests cover the new or changed behavior. Release Notes: - Fixed remote projects (SSH, WSL) becoming unresponsive when loading an agent thread that refers to many files outside the project. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
|
03efd49220 |
ui: Keep project action shortcuts inside their buttons (#65531)
## Summary This is my first PR; I'm still learning Rust. - We made long project names show dots at the end when they don't fit, so the shortcut stays visible. - We made the shortcut go underneath Open Project when the sidebar is too narrow, and the button grows so everything fits. - We added an option for a button to grow to fit its contents, and Open Project uses it when the shortcut wraps. AI helped with the code and testing. Closes #65220. ### Screenshots Welcome before: Zed 1.23.3. The other screenshots are local development builds on macOS. | | Before | After | | --- | --- | --- | | Welcome recent project |  |  | | Empty Project Panel |  |  | ## Testing - Tested with Codex on macOS 27.0.1 in light and dark themes, using the default font and Menlo at 16px. - Clicking and keyboard shortcuts worked. The build, formatting, Zed's lint checks, and six focused tests passed. - Linux and Windows still need testing. To try it: 1. Open a folder with a long name, then a new window. Hover over the recent project. 2. Set the sidebar to 180px wide, use Menlo at 16px, and bind Ctrl-K Ctrl-O to Open Project. Hover over the button and check that both lines fit. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. Release Notes: - Fixed long recent project names and the Open Project shortcut overflowing on the Welcome screen and in narrow panels. --------- Co-authored-by: Smit Barmase <heysmitbarmase@gmail.com> |
||
|
|
6744a21e2d |
agent_servers: Route ACP subagent associations to child threads (#65976)
ACP agents announce subagent sessions with `subagent_update` on the parent's stream. Zed dropped these: v2 left them unhandled and v1 ignored them. On v2, where subagent support is part of the baseline and only negotiated with `acp-beta`, a subagent's updates were discarded and its permission requests failed with "unknown session". - **Routing.** The first `subagent_update` for a session creates an agent-owned child `AcpThread` (no prompting, no working directories, action log linked to the parent's, as native subagents do), registers it for routing on the connection, and hands it to the parent's subagent registry. The child's own updates, permission, elicitation, fs and terminal requests then reach it. One helper, `apply_subagent_update`, serves v1 and v2. - **Association.** Title, description and cancel capability are stored on the child (`acp_thread::subagents::SubagentAssociation`) with the RFD's patch semantics: omitted is unchanged, `null` clears, capabilities are replaced whole. A session ID the parent does not already own is rejected if the connection routes it, so an update cannot reparent a session, refer to the parent itself, or form a cycle. - **Activity.** On v2 the subagent's own `state_update` drives its activity and the parent's mirrored copy is ignored. On v1 the association's state snapshot is the only source, and `null` leaves activity unconfirmed. - **v1 stays off.** v1 applies `subagent_update` only if Zed advertised the `subagents` client capability, decided from the same `client_capabilities_for_agent` value it sends. Zed does not advertise it yet (a test guards that), so v1 behavior is unchanged; a later PR offers it behind `acp-beta`. - **No closing subagents.** Releasing a child never sends `session/close`, and v2's `close_session` refuses ACP subagents; the agent releases them with their parent. - **UI.** A subagent's permission card in the parent labels it with the parent's title for it, then its own title. Inline subagent cards and navigation into subagent threads come in a later PR, so today the visible change on v2 is that subagent permission requests appear in the parent and can be answered. Release Notes: - N/A |
||
|
|
9b9e7b814a |
xtask: Forbid legacy Objective-C dependencies (#65935)
Reject objc, cocoa, and related legacy wrappers in workspace packages and unused workspace dependency declarations. Cover aliases, optional and target-specific dependencies, build and dev dependencies, and permitted objc2 replacements. Release Notes: - N/A |
||
|
|
2d6c664dc2 |
gpui_macos: Cache the main-thread marker per thread (#65934)
Reuse a thread-local `Option<MainThreadMarker>` for platform creation, display queries, and display-link assertions. Preserve background-thread rejection and cache both successful and unsuccessful checks. An instrumented standalone check observed one pthread_main_np call per thread across repeated main-thread and worker-thread lookups; the uncached implementation failed with 1000 calls. Release Notes: - [GPUI] Cache macOS main-thread checks per thread. |
||
|
|
5d47f83542 |
Allow opening README files by default (#65389)
Closes https://github.com/zed-industries/zed/discussions/38228 https://github.com/user-attachments/assets/4f80a4a8-8e4a-499b-b51f-4d389fe1f781 Release Notes: - Added a way to open README files by default, use `"on_new_window": "readme",` to configure |
||
|
|
b7d50b95e8 |
Warn only for explicitly requested fonts (#65968)
Follow-up to https://github.com/zed-industries/zed/pull/65852 Release Notes: - [GPUI] N/A |
||
|
|
10cfb45944 |
gpui_macos: Migrate window and appearance bindings to objc2 (#65933)
Migrate window classes, callbacks, native geometry, ownership, and appearance handling to objc2. Remove the temporary legacy adapters and the now-unused objc and cocoa dependencies. Release Notes: - [GPUI] Migrate macOS window and appearance bindings to objc2 |
||
|
|
faf8f0b698 |
language_models: Log and record model-fetch failures instead of discarding them (#64171)
# Objective Addresses part of #58102. When a model list fetch fails (e.g., due to an invalid or expired API key), Zed currently swallows the error without logging it anywhere. Root cause: Across six provider implementations (anthropic, ollama, lmstudio, llama_cpp, open_router, and vercel_ai_gateway), restart_fetch_models_task stores the fetch Task in a struct field but never inspects its Result. Whenever a new fetch is triggered, .replace() simply drops the previous task and silently discards whatever error occurred. ## Solution Added `last_fetch_models_error: Option<String>` to each provider struct to track the failure. In `restart_fetch_models_task`, the fetch is now wrapped in an outer task that logs the failure via `util::ResultExt::log_err()` and updates that field. Wrapping the task rather than calling `.detach_and_log_err()` keeps task cancellation intact: dropping the outer task when a new fetch starts still cancels the in-flight request, as expected. `fetch_models` itself is unchanged, since callers like `authenticate` in `lmstudio.rs` and `llama_cpp.rs` match directly on its `Result`. This PR is deliberately limited to logging and recording the error. Surfacing it in the UI would be better tackled in a dedicated follow-up. That keeps this PR focused on the backend plumbing across the six providers. ## Testing Added `restart_fetch_models_task_records_error_on_failure` across all six providers, using `FakeHttpClient` to confirm `last_fetch_models_error` gets populated on failure. Added `restart_fetch_models_task_cancels_the_previous_fetch` to `anthropic.rs` (using channel-gated fake responses) to verify that an in-flight request is cancelled and cannot overwrite a newer fetch. All checks pass cleanly: `cargo test -p language_models` (97 passed), `cargo clippy -p language_models -- -D warnings`, and `cargo fmt --check -p language_models`. UI testing: N/A (no UI changes in this PR; all verification is headless). ## Self-Review Checklist - [x] I've reviewed my own diff for quality, security, and reliability - [x] Unsafe blocks (if any) have justifying comments - [ ] The content adheres to Zed's UI standards (n/a — no UI changes) - [x] Tests cover the new/changed behavior - [x] Performance impact has been considered and is acceptable (single Option<String> per provider, zero happy-path allocation overhead) --- Release Notes: - Fixed: failed model list fetches (such as from invalid or expired API keys) are now logged and recorded instead of silently dropped across Anthropic, Ollama, LM Studio, llama.cpp, OpenRouter, and Vercel AI Gateway. Co-authored-by: Kirill Bulatov <kirill@zed.dev> |
||
|
|
d86aa6e6b9 |
gpui_macos: Migrate application and menu bindings to objc2 (#65932)
Migrate application delegates, menus, platform services, and retained AppKit objects while preserving checked ivar access. Kept legacy window string allocation and an appearance pointer cast until the window migration (next PR). Release Notes: - [GPUI] Migrate macOS application and menu bindings to objc2 --------- Co-authored-by: Piotr Osiewicz <24362066+osiewicz@users.noreply.github.com> |
||
|
|
3389d87577 |
Cache resolved fonts (#65852)
Closes https://github.com/zed-industries/zed/issues/65777 Release Notes: - [GPUI] Fixed font lookup causing performance issues |
||
|
|
0eb085876e |
Use recent commits to improve message generation (#61528)
## Summary Commit message generation now includes the three most recent repository commit messages as style examples alongside the diff. ## Problem The generator previously had no visibility into the repository's recent commit conventions. As a result, generated messages could drift from the project's established style, including its use of scopes, bodies, and release-note sections. ## Benefits - Keeps generated commit messages aligned with the repository's current conventions. - Preserves full recent commit messages so the model can learn both subject and body formatting. - Handles repositories with fewer than three commits or unavailable history without blocking generation. - Starts history lookups alongside diff collection to avoid adding unnecessary serial latency. ## Testing - `git diff --check` passed. - `cargo test -p git_ui test_commit_message_prompt --lib` could not run because `cargo` is not installed in the environment. - `cargo fmt --all -- --check` and `./script/clippy -p git_ui` could not run for the same environment limitation. Release Notes: - Improved AI commit message generation by using recent repository commit history for style consistency. Last 3 commit messages are now sent out to the message-generating model as reference for style. --------- Co-authored-by: Piotr Osiewicz <24362066+osiewicz@users.noreply.github.com> Co-authored-by: zed-zippy[bot] <234243425+zed-zippy[bot]@users.noreply.github.com> |
||
|
|
0a39cdb4f3 |
gpui: Fix window_positioning example on non-primary displays (#65917)
## Summary The `window_positioning` example opens nine popups on each display, but five of them (Top Left, Top Center, Left Center, Right Center and Bottom Center) work out their position from `(0, 0)` or from the display's size instead of from the display's bounds. That works on the primary display, but on any other display those spots end up off the display. On Windows, gpui then falls back to the default bounds, so those popups open display-sized instead of where they should be. This changes those five to use `screen.bounds()` (`origin`, `left()`, `right()`, `top()`, `bottom()`), like the other four already do. Only the example is changed. Screenshots of both displays, primary on the left. Both builds have #65002 applied, since without it the popups on the 100% display come out the wrong size, which gets in the way of the comparison. **Before** <img width="1000" alt="zed-window-positioning-before" src="https://github.com/user-attachments/assets/4707993d-efb9-4d15-95ac-8433c61d4a0f" /> **After** <img width="1000" alt="zed-window-positioning-after" src="https://github.com/user-attachments/assets/ebbaea7c-ef82-4512-9522-921e9fe5fd99" /> In the before shot the second display is mostly black because Right Center opened display-sized on top of the others. Its label shows it asked for `940, 412.5` and got `2560, -4` at 1440x900. ## Testing Windows 11 (build 26300), a 2560x1600 primary display at 150% and a 1440x900 display at 100% to its right. I ran the example before and after and checked where each popup ended up on its display: - Displays aligned at the top: 6 of 9 popups in place on the second display before, 9 of 9 after. - Second display moved 696px down: 4 of 9 before (and Bottom Center sits partly above the display), 9 of 9 after. - Primary display: 9 of 9 in every run. `cargo fmt` and `cargo clippy -p gpui --example window_positioning -- --deny warnings` pass. I haven't tried macOS or Linux, but nothing in the change is platform-specific. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). (No UI changes in Zed, only a gpui example.) - [ ] Tests cover the new or changed behavior. (The examples have no automated tests. I tested it by hand on two displays, see Testing.) I used Claude to help investigate this and write the change. I've reviewed it and understand it. Release Notes: - [GPUI] N/A |
||
|
|
87c92e8c09 |
Add .agents folder to .gitignore (#57857)
This PR adds .agents to zed's .gitignore file to allow people to have to have local only skills We can still manually add skills to Zed when people want to share them. This should make it easier for people have their own custom skills while still being able to run `git add .` Self-Review Checklist: - [x] I've reviewed my own diff for quality, security, and reliability - [x] Unsafe blocks (if any) have justifying comments - [x] The content is consistent with the [UI/UX checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) - [x] Tests cover the new/changed behavior - [x] Performance impact has been considered and is acceptable Release Notes: - N/A --------- Co-authored-by: MrSubidubi <finn@zed.dev> |
||
|
|
130f1e92a4 |
Update dependencies to address security vulnerabilities (#65956)
## Summary Update Zed's TLS certificate dependencies, managed Node runtime, and Wasmtime stack to address the eMudhra intermediate compromise and recent Wasmtime sandbox vulnerabilities. - Update `rustls-webpki` to 0.103.16, `webpki-roots` to 1.0.10 and its legacy wrapper to 0.26.11, and `webpki-root-certs` to 1.0.10. SQLx's legacy wrapper resolves to the patched 1.0.10 root data. - Update managed Node from 24.21.0 to 24.22.1. - Update Wasmtime, WASI, and Tree-sitter's Wasmtime C API from 48.0.1 to 48.0.7, with aligned Cranelift 0.135.7 and required Wasm tooling patches. Keep the existing Wasmtime feature set. ## Testing Local platform: macOS arm64, Rust 1.98.1. - `cargo build --locked -p zed --bin zed`: passed. Launched the new binary with isolated user data and a temporary Rust file; the process remained running. A user-provided screenshot confirms `app-smoke.rs` opened and rendered with Rust syntax highlighting. Interactive GUI checks were not completed: the UI automation tool failed to initialize and macOS denied the Apple Events fallback. This PR remains a draft for human review and any further manual validation. - `ZED_HEADLESS=1 cargo test --locked -p extension_host -p node_runtime -p reqwest_client --lib`: passed. The extension-host integration test builds, installs, and runs the real test extension, including relative/absolute filesystem operations and process execution. ```text extension_host: 46 passed; 0 failed node_runtime: 24 passed; 0 failed reqwest_client: 4 passed; 0 failed ``` - `cargo check --locked -p extension_host -p extension_cli -p language`, `./script/clippy --locked -p node_runtime -p reqwest_client`, formatting, and diff whitespace checks passed. The lint wrapper's optional cargo-shear/typos/buf checks were not run because cargo-shear was unavailable. - Isolated native aarch64 checks passed: oversized branch tables rejected at 65,521 and 262,146 targets; a table at the supported 65,520 limit executes correctly; component instantiation and the pinned Tree-sitter C API WasmStore work. WASI read/write/rename and internal symlinks work with Zed's dual ReadWrite preopens; the trailing-slash outside symlink rename and parent traversal are rejected without modifying the outside sentinel. - Isolated TLS checks passed: WebPKI 0.103.16 rejects all 12 captured incident certificate chains with `CertDistrusted`, including when supplied the old roots. The patched root bundles omit the affected root. The official native Node archive matches its published SHA-256; Node rejects all 12 incident hosts with `CERT_REVOKED` and accepts an ordinary verified TLS control request. - The standalone probes are local validation artifacts, not added repository tests. Reviewers can rerun the Cargo commands above; for manual validation, launch the build, open a source file, and install/run a Wasm extension and a Wasm grammar. Linux/Windows runtime testing and a full workspace test run were not performed. The debug app build emits a large unwind-table linker warning and a future-compatibility warning for `block`. ## Self-review The dependency diff was reviewed for scope, version alignment, feature changes, and sandbox compatibility. Two incidental dependency-link changes selected by Cargo were removed; only the intended security updates remain. README's review notice is retained because repository rules reserve its removal for the human author. - [x] I've reviewed my diff for quality, security, reliability, and performance. (Human author review pending.) - [ ] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). (N/A; no UI changes.) N/A - [x] Tests cover the new or changed behavior. ## Suggested .rules additions When checking a Cargo.lock update for scope, compare dependency lists for packages whose own versions did not change; Cargo can reselect their already-locked transitive versions. Release Notes: - Updated dependencies to address recent software ecosystem vulnerabilities. |
||
|
|
b488c4e814 |
gpui: Update Taffy 0.13 to 0.15 (#65950)
Updates Taffy from 0.13 to 0.15. Upstream reports roughly 60% faster grid layout and 15% faster flexbox layout, alongside layout caching improvements and fixes to sizing, alignment, and absolute positioning. This adapts GPUI to the updated measurement and alignment APIs while retaining its explicit relative-positioning default; performance improvements have not been benchmarked in Zed. Validation: All 469 existing GPUI library tests passed with `corgi test -p gpui --lib -F gpui_platform/runtime_shaders`; formatting and diff checks passed. Runtime shaders were enabled to bypass a local Metal compiler failure. Release Notes: - [Gpui] Update taffy version |
||
|
|
544780422a |
Add an HTTP fallback for the Ubuntu toolchain PPA (#65944)
## Summary Add an HTTP fallback for the Ubuntu toolchain PPA used by `script/linux` on Ubuntu 20.04. The [Linux ARM64 release job](https://github.com/zed-industries/zed/actions/runs/39623264964/job/119388744726) failed because Launchpad's HTTPS frontend returned HTTP 503 for package downloads. The official HTTP endpoints served the same packages successfully. Use APT's native `mirror+file` transport with explicit priorities: prefer Canonical's HTTPS endpoint, and use Canonical's HTTP endpoint only when a fetch fails. This handles both repository indexes and package archives without custom retry logic. Other repositories and distro setup paths are unchanged. ## Security tradeoff This is an accepted availability workaround, not a security-equivalent replacement for HTTPS. The signing key is still fetched over HTTPS, and APT signature and package-hash verification remain enabled. HTTP fallback permits network observers to see requests and attackers to replay or freeze previously signed repository contents. This PPA has no `Valid-Until` field. It does not permit arbitrary package modification without a verification failure. ## Testing - `bash -n script/linux` — passed. - `nix shell nixpkgs#shellcheck -c ./script/shellcheck-scripts error` — passed. - Temporary integration harness using Ubuntu 20.04 / APT 2.0.10 and a signed local repository, exercising the mirror-generation commands extracted from `script/linux`: - Healthy HTTPS serves all requests without contacting HTTP. - HTTP 503 on the primary repository index falls back successfully. - HTTP 503 on a package download falls back successfully, even after indexes were fetched over HTTPS. - Both endpoints failing produces a nonzero exit. - Tampered signed metadata is rejected. - Tampered package contents are rejected. - Downloaded all seven failing ARM64 packages across Canonical's official HTTP endpoints and verified their hashes against repository metadata; verified the signed `InRelease` and package-index hash. - No full dependency installation or CI rerun performed. The temporary integration harness is not committed. Release Notes: - N/A |
||
|
|
42c4caa474 |
remote: Stop URL-encoding SSH usernames (#64744)
The username was percent-encoded when building the SSH destination, a leftover from when the destination was an `ssh://` URL (#25314). Since the destination is now passed as a plain argument, the encoding turned domain usernames like `DOMAIN\user` into `DOMAIN%5Cuser`, and authentication failed. Pass the username verbatim instead. So that a username beginning with `-` can't be parsed as an option, `ssh` now receives it via `-l <username> -- <host>` rather than a `username@host` string, and `scp`/`sftp` (which have no username option, and forward it to `ssh` via `-l` themselves) get `--` before their operands. Options that used to follow the destination (`-T`, and `-t` for the Windows master process) are moved before `--` so they aren't sent as part of the remote command. Fixes #63709 Closes #63722 Assisted-by: LLM NOTE: This is a continuation of #63722. I chose to leave Shaik Shahid Afridi as the primary author, because I'm just resolving the review comments from that PR. In responding to the review comments, we also improve SSH/SCP for usernames starting with `-`, because we now use `-l` for SSH and `--` for SFTP/SCP, though I'm not sure if OpenSSH would accept a username starting with `-` anyways. Nevertheless, an improvement over what came before. # Objective When attempting to connect to a remote with a backslash in the username (or any other character that results in URL encoding), Zed would mangle the resulting SSH command and the remote would not be usable. ## Solution - Describe the solution used to achieve the objective above. URL encoding in the SSH field is an artifact of when `ssh://` was used, and is no longer needed. ## Testing Tested by running Zed and attempting to connect to a remote with a backslash in the user field (a Windows domain account). SSH succeeded, but then I ran into a separate bug with `%PROCESSOR_ARCHITECTURE%` not being expanded. I'll be looking into that in a separate PR. ```sh 2026-09-24T16:09:24-05:00 INFO [remote::transport::ssh] Remote is windows: true 2026-09-24T16:09:38-05:00 ERROR [remote::transport] $SHELL is not set, falling back to cmd.exe 2026-09-24T16:09:38-05:00 INFO [remote::transport::ssh] Remote shell discovered: cmd.exe 2026-09-24T16:09:39-05:00 ERROR [workspace::notifications] Prebuilt remote servers are not yet available for windows-%PROCESSOR_ARCHITECTURE%. See https://zed.dev/docs/remote-development ``` Tested on Linux x86_64, attempting to connect to a Windows x86_64 machine. ## Self-Review Checklist: - [x] I've reviewed my own diff for quality, security, and reliability - [ ] Unsafe blocks (if any) have justifying comments **(N/A)** - [ ] The content adheres to Zed's UI standards ([UX/UI](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) and [icon](https://github.com/zed-industries/zed/blob/main/crates/icons/README.md) guidelines) **(N/A)** - [x] Tests cover the new/changed behavior - [x] Performance impact has been considered and is acceptable ## Showcase **NOTE:** I mentioned it above, but just in case it was missed, there is a separate bug with Zed not expanding `%PROCESSOR_ARCHITECTURE%`. I'll probably look into that next, since I'm trying to get remote development set up for my Windows machine. [Screencast_20260924_160804.webm](https://github.com/user-attachments/assets/ce6c77fb-eea5-4cda-944e-50539bece61d) --- Release Notes: - Fixed an issue where SSH usernames containing `/` or `\` were percent-encoded, preventing successful connections. Co-authored-by: Shaik Shahid Afridi <81465960+Yooosless@users.noreply.github.com> Co-authored-by: Kirill Bulatov <kirill@zed.dev> |
||
|
|
63089a6ebe |
worktree: Clean up watches when symlinks disappear during scanning (#65939)
## Summary Fix a watch leak introduced by #61372: deleting an external symlink between watch registration and directory enumeration left its canonical target watched. Record cleanup and event-path mappings when establishing the watch, before enumeration can fail. Add a deterministic FakeFs regression. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. Release Notes: - N/A |
||
|
|
cb593458e1 |
project: Match document filter patterns in dynamic registrations (#65252)
## Summary Follow-up to #59243, which made dynamically registered capabilities honor the `language` and `scheme` of their document selectors, but left `pattern` matching fail-open. A registration such as `{ "language": "toml", "pattern": "**/Project.toml" }` therefore still applied to every TOML document the server had opened, and a pattern-only filter applied to every document. This matches `pattern` against the document's path: - String patterns are matched against the absolute path. As in the LSP glob syntax, `*` and `?` do not match path separators. - Relative patterns are matched against the path relative to their base URI, and Zed now advertises the `textDocument.filters.relativePatternSupport` client capability. - Documents without a file system path never match a `pattern` filter. Contexts that stand for a whole language, like the semantic token legend lookup, and guest buffers whose path is unknown keep not evaluating patterns. - Registrations whose patterns are not valid globs are rejected when they are registered. LSP 3.18 changed `TextDocumentFilter.pattern` from a string to a `GlobPattern`, so that it can also be a relative pattern. ~~This PR depends on zed-industries/lsp-types#16, which updates the type and adds the client capability. I'd like to merge this together with that upstream change: `Cargo.toml` currently points at the commit of zed-industries/lsp-types#16 on my fork, and I will switch it to the `zed-industries/lsp-types` revision once that PR is merged.~~ **Update:** zed-industries/lsp-types#16 has been merged, and `Cargo.toml` now points at its `zed-industries/lsp-types` revision (zed-industries/lsp-types@82d7fc66eb). ## Testing - Added unit tests for matching string patterns, wildcards across path separators, and relative patterns. - Added an integration test that a pattern-scoped completion registration only applies to matching buffers, and that a registration with an invalid pattern is rejected. - Updated `test_dynamic_formatting_registration_with_pattern_only_selector`, which asserted the fail-open behavior left by #59243. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. Release Notes: - Fixed language servers' dynamically registered capabilities applying to files outside of the `pattern` of their document selector. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
|
59dd60d4dd |
Refactor ExtensionLanguageProxy to provide only ExtensionLanguageProxy::update_languages (#65381)
## Summary Further preparation for https://github.com/zed-industries/zed/pull/65047 - this PR changes the trait so thar we now register and unregister languages through a single call into the trait. This allows us to easier manage state depending on this - we previously always removed and readded everything, with this now in place, we instead get a diff that let's us know what changed and thus needs updating. That will be used in the linked PR as well. ## Testing Added some tests (actually most of the diff), other behavior is still the same, and also gave it a manual test run. Tests caugtht some edge cases with token rules previously unhandled which this PR now also takes care of. ## Self-review - [X] I've reviewed my diff for quality, security, reliability, and performance. - [N/A] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [X] Tests cover the new or changed behavior. Release Notes: - N/A --------- Co-authored-by: zed-zippy[bot] <234243425+zed-zippy[bot]@users.noreply.github.com> |
||
|
|
cd1255fba1 |
Automate cleanup of non-responsive PRs (#65930)
Ping and (after a grace period) close the stale community PRs where changes were requested but not addressed. Ping and (after a grace period) close the community PRs with no self-review. ## Testing Ran the dry-run locally and verified every PR listed. Release Notes: - N/A |
||
|
|
2afec34882 |
Prevent existing tabs on move from emitting AddItem on reordering (#65795)
## Summary Moving a pane shouldn't issue a `AddItem` trigger event on pane's `add_item_inner` function. Treating tab reordering as a new event would cause listeners like markdown auto preview mode to be triggered (if enabled) as well which will activate preview mode even if we're currently in edit mode for a markdown file. Closes #65545 ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [ ] Tests cover the new or changed behavior. Release Notes: - Fixed reordering a markdown editor tab switching back to markdown preview when "open markdown files in preview" is enabled |
||
|
|
d27ad54dd6 |
Fix on-type formatting using a stale position when autoindent is pending (#65769)
## Summary Sister to Shopify/ruby-lsp#4242. Pressing Enter triggers `textDocument/onTypeFormatting` from inside the newline transaction, before autoindent has been applied. The request position was converted to a point at that moment, so language servers received column 0 of the new line even though the cursor ends up at the indented column. The left-biased cursor pins from #61823 were captured at the same moment, so the autoindent whitespace prevented them from being restored. With Ruby LSP this inserted `end` in front of the body's indentation and left the cursor after `end`. That is why the Ruby extension disables on-type formatting (zed-extensions/ruby#38). This PR: - Resolves the on-type formatting position after pending autoindent is applied. - Capture the cursor pins after pending autoindent is applied. ## Testing - `cargo test -p editor --lib on_type_format` (7 passed; the 2 new tests fail on `main`) - `cargo test -p collab --features test-support --test collab_tests on_input_format` (2 passed) - `./script/clippy -p editor -p project` - Manually with Ruby LSP: `def foo⏎`, `do |x|⏎`, heredocs, `do |` and `"#{` ### Proof https://github.com/user-attachments/assets/43ff3307-4bd7-43f5-a65d-d18eca0b5d2a ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. ## References - zed-extensions/ruby#38 - Follow-up to #61823 ## Suggested .rules additions For `crates/editor/.rules`: code that runs inside `Editor::transact` (for example from `newline`) sees the buffer before autoindent is applied. Wait for `Buffer::wait_for_autoindent_applied()` before reading cursor positions or anchoring to them. Release Notes: - Fixed on-type formatting after pressing Enter being applied before the new line's indentation, which left the cursor after text inserted by language servers such as Ruby LSP. |
||
|
|
f41ac64812 |
agent: Parse thread markdown lazily, on first render (#65470)
## Summary Opening an agent thread previously materialized parsed Markdown for message and tool-output blocks before those blocks were laid out, including blocks far off-screen. Parsed Markdown can retain events, per-code-block syntax highlights, and rasterized Mermaid diagrams. The original debug-build profiling with two large threads open attributed approximately 295 MiB to `markdown`, with individual assistant or tool blocks accounting for tens of MiB. The profiling figures below are from the original measurements and were not rerun during the final review. The thread view uses a virtualized `gpui::list` (`ListState`), but layout also includes overdraw and can include focused off-screen rows. Previously, parsing started at entity creation in `acp_thread`, independently of list layout. With `MarkdownOptions::defer_parsing`, `Markdown` records the source without parsing until one of these occurs: - `MarkdownElement::request_layout` lays out the block. - `Markdown::ensure_parsed(cx)` is called explicitly. While deferred, source updates through `append` and `reset` do not start parsing. Once parsing starts, later source updates parse normally, including when the block is off-screen. This is one-time lazy activation, not eviction of previously parsed blocks. `acp_thread` enables deferral for message content created through `ContentBlock::create_markdown`, including assistant and tool messages and compaction summaries, and for tool-call raw input and output. Tool-call labels and error text remain eager. The numbered `read_file` output bypasses `MarkdownElement` to draw its own gutter. Its `RenderOnce` component now calls `ensure_parsed` before reading the code-block language for highlighting. Raw source remains available before parsing for search, copy/export, and content-visibility checks. Restoring a thread applies its saved scroll position before deferred Markdown has finished measuring. `ListState::set_item_measurement_readiness` lets the thread report pending Markdown parsing and Mermaid rendering, including compaction summaries and errors. The list preserves scroll intent while measurements remain provisional. Leading overdraw stops at the first unready preceding row rather than traversing the entire zero-height prefix. Search autoscroll also waits for parsing and Mermaid rendering to finish. ## Original memory measurements The original before/after heap profiles used uncapped dumps of the same view: two large threads and one expanded tool card. | | Before | After | | ---------------------------------------- | --------: | --------: | | `markdown` category | 273.7 MiB | 109.5 MiB | | Parsed events (`start_background_parse`) | 213.0 MiB | 11.0 MiB | Deferred blocks that have not been laid out or explicitly activated do not retain a parsed representation. Raw source remains available, and blocks that have already been parsed are not evicted merely because they leave the viewport. Companion PR: #65444. ## Testing - Agent UI library suite: 519 passed. - Three scroll-restoration regressions across 20 scheduler seeds each: 60 successful runs. - GPUI list tests: 36 passed. - Focused Markdown/Mermaid tests: 21 passed. - ACP compaction regression: passed. - Five regression assertions fail with the corresponding fixes removed. - Two adapted controls fail on main with production code unchanged. - Scoped `rustfmt` and `git diff --check` passed. Validation used isolated Cargo build artifacts and `INPUTRC=/dev/null` for the agent-UI suite. Interactive visual behavior was not manually validated during this review. Release Notes: - Improved agent panel memory usage by parsing thread message Markdown on first layout instead of at thread load. - [GPUI] Added `ListState::set_item_measurement_readiness` to preserve scroll intent while asynchronously measured list items settle. --------- Co-authored-by: Kirill Bulatov <kirill@zed.dev> Co-authored-by: Kirill Bulatov <mail4score@gmail.com> |
||
|
|
9e65fc7682 |
agent: Materialize tool-call diffs lazily, on first view (#65444)
## Summary Restoring a thread materialized the diff of every edit in its history, whether or not the card had ever been shown. On long, tool-heavy threads that meant multi-GiB memory spikes just from opening a thread. Two eager paths on restore: - `EntryViewState::sync_entry` built a diff `Editor` and a `TerminalView` for every tool call in the history. - `acp_thread::Diff::finalized` created a `Buffer::local(new_text)` holding the whole post-edit file, set the language from the file path, and awaited `parsing_idle()`, retaining the parsed buffer in `FinalizedDiff`. Profiling a real 44-thread store (debug build, isolated `LD_PRELOAD` heap profiler instrumenting only the `zed` process) attributed ~1.26 GiB to tree-sitter syntax trees and ~3.0 GiB anon to opening a thread, with the dominant stacks under `Buffer::reparse` / `SyntaxSnapshot::reparse_with_ranges` reached from `Diff::finalized`. The diff-editor and terminal widgets were a negligible part of it. Diffs are now built on first view: - `EntryViewState::sync_entry` no longer builds diff and terminal content views for tool calls whose content is not visible. The history-restore loop calls `sync_entry_initial`, and expanding a card re-syncs that entry to build its views. - `acp_thread::Diff::finalized` stores only the diff source (`path`, `old_text`, `new_text`). The buffer, rope, multibuffer and tree-sitter parse are built by `Diff::resolve`, called from `AcpThread::resolve_tool_call_diffs(index, cx)` when the card first becomes visible. `resolve` is idempotent, and the accessors reachable before resolution (`to_markdown`, `has_revealed_range`) handle the unresolved case. Same isolated profile, same 44 threads, `MPROF_MIN=4096 MPROF_SAMPLE=64`: | | before | after | |---|---:|---:| | `Pss_Anon` after load | ~3.0 GiB | 478 MiB | | `Rss` | 3.4 GiB | 997 MiB | | tracked live heap (est.) | 1933.6 MiB | 184.7 MiB | | syntax / tree-sitter | 1262.8 MiB | 7.1 MiB | Opening a second thread kept tree-sitter at ~6 MiB, and expanding one card built only the diff for that card (~12 MiB), so the work is now proportional to what is shown. Companion PR: #65470 parses thread markdown lazily, on first render. With both changes the same profile goes from 3003 to 1199 MiB `VmRSS`, and anonymous memory from 2724 to 934 MiB.  Still eager on restore, and left for follow-ups: markdown parsing of assistant messages (~245 MiB over two threads), `serde_json` deserialize and `Value::clone` of the thread payload (~276 MiB), mermaid rasterization (~59 MiB), and the same lazy treatment for `DiffPatch` hunks. Related reports: #65465, #62133, #63720, #59402, #60793, #57478, #57126, #61026, #50889. ## Testing - `cargo check -p agent_ui` - `cargo test -p acp_thread` - `cargo test -p agent_ui` Release Notes: - Improved memory usage in the agent panel: tool-call diffs are built and parsed only when a card is first viewed, instead of for every edit in a thread's history. --------- Co-authored-by: Kirill Bulatov <kirill@zed.dev> |
||
|
|
3dbf6a8101 |
Fix npm-backed language server setup in pnpm and Yarn projects (#65280)
## Summary Give directory-less npm commands a Zed-owned working directory and explicit prefix, preventing project devEngines requirements from blocking language server package queries. Apply the directory consistently to managed and system runtimes while preserving explicit install targets. Cover command construction, child working directories, and propagation of directory-creation errors for both runtimes. Fixes #65279 ## Testing Tested locally and confirmed language server setup works in Yarn and pnpm projects. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. Release Notes: - Fixed language server setup failing with `EBADDEVENGINES` in Yarn and pnpm projects. --------- Co-authored-by: Codex <noreply@openai.com> |
||
|
|
1563b405f9 |
project: Honor document selectors of didChange and didSave registrations (#65253)
## Summary #59243 made dynamically registered capabilities honor their document selectors, but `textDocument/didChange` and `textDocument/didSave` registrations still overwrote the server-wide sync options and ignored their selectors, while `didSave` was sent for every saved file of the worktree to every server whose options asked for saves. These registrations are now kept per registration, with their selectors: - The sync kind of a document's changes and whether its saves are notified, with or without text, come from the latest registration matching the document, falling back to the static options. - The merged capabilities keep mirroring the latest registrations, and unregistering now restores the static options instead of clearing them. - A `didSave` registration without options is now honored instead of ignored. Static save options now only cover the documents a server was opened with, as they apply to the documents the client selects for the server, while `didSave` registrations may select any document of the worktree. Registrations without a selector stand for the same documents as the static options. This removes the broadcast of every save to every server, which rust-analyzer no longer needs to pick up manifest changes in Zed: with dynamic registration it declares no static save options, and registers `didSave` for `**/*.rs` and `**/Cargo.toml` to reload its workspace when a manifest is saved. Manifest saves keep reaching it either way: while `pattern` matching remains fail-open, its filters match every document, and once #65252 lands they select exactly these files. Documents without a language no longer match every registration, and only match filters without a `language`. ## Testing - Added integration tests that `didChange` and `didSave` registrations only apply to documents matching their selectors, fall back to the static options otherwise, and restore them when unregistered. - Added an integration test that static save options and registrations without a selector do not cover documents not opened in the server, that a `language` filter does not select a document without a language, and that a registration can select such a document. - Updated `test_managing_language_servers`, which asserted that saving a file notifies every server of the worktree. ## Self-review - [x] I've reviewed my diff for quality, security, reliability, and performance. - [x] UI changes follow the [checklist](https://zed.dev/docs/development/ui-checklist). - [x] Tests cover the new or changed behavior. Release Notes: - Fixed language servers receiving change and save notifications that do not follow their dynamic registrations, including save notifications for unrelated files of the project. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> |
||
|
|
11262f9256 |
project_panel: Explain when a move fails because the destination already exists (#64344)
# Objective - Fixes #64128 dragging a file onto a folder that already contains a file with the same name showed only Error: renaming "/a/1.json into "1.json". The reason why was buried inside the error because the pop-up prints only the outermost layer. ## Solution undo.rs: -Extracted the existing already_exists check into a module level pub(crate) function which walks the error chain for io::ErrorKind::AlreadyExists with a fallback that looks for the text "already exists." Undo's rename now calls that function so it behaves exactly like before. - Made project_path_display() a pub(crate) fn so the project panel can format names the same way undo does. project_panel.rs: - move_worktree_entry() now hands back the destination path along with the rename task instead of discarding it. - drag_onto() builds (from, to) names from old_paths and the returned destination. This hapens while the project can still be read and before anything moves. - explain_move_error(err, names) -> anyhow::Error when names are available and the error is "already exists," it adds the clear message with err.context(...), otherwise it returns the error unchanged. - drag_onto() - both result loops (normal and folded folders) - each now destructures three items and calls .map_err(|err| explain_move_error(err, names.as_ref()) just before .notify_workspace_async_err() ## Testing - (#[gpui::test]) test_drag_onto_existing_name_shows_clear_error drags a/1.json onto a root that already has 1.json. It checks that both files keep their original contents and that exactly one notification appears. - (#[test]) test_explain_move_error covers the three cases 1. an already-exists error gets the clear message. 2. a different error passes through untouched. 3. with no names, the error passes through untouched. - Both new tests pass. - The unit test fails when detection is broken on purpose by returning false in the new pub(crate) fn already_exists_error, so it does guard the message logic. ## Self-Review Checklist: - [x] I've reviewed my own diff for quality, security, and reliability - [x] Unsafe blocks (if any) have justifying comments - [x] The content adheres to Zed's UI standards ([UX/UI](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) and [icon](https://github.com/zed-industries/zed/blob/main/crates/icons/README.md) guidelines) - [x] Tests cover the new/changed behavior - [x] Performance impact has been considered and is acceptable ## Showcase <details> <summary>Click to view showcase</summary> <img width="495" height="150" alt="before_zed_project_panel_name_error" src="https://github.com/user-attachments/assets/735eb482-f4b9-4502-b810-782180c3b40c" /> <img width="495" height="150" alt="after_zed_project_panel_name" src="https://github.com/user-attachments/assets/e7e9c30f-e171-4ce7-a3fd-c0f2690c08f5" /> <img width="217" height="186" alt="zed_test_panel_shot" src="https://github.com/user-attachments/assets/8d76e18a-9695-4f4a-8231-785ae140c73e" /> </details> --- Release Notes: - Improved the error message in the project panel when moving a file or folder onto a location where one with the same name already exists --------- Co-authored-by: Kirill Bulatov <mail4score@gmail.com> |
||
|
|
e83f1c9495 |
gpui_macos: Migrate clipboard bindings to objc2 (#65713)
Use typed pasteboards and retained Foundation objects while preserving clipboard formats and metadata. Cover embedded NUL and Unicode text and autorelease lifetimes. Release Notes: - [GPUI] Migrated macOS clipboard bindings to objc2 |
||
|
|
cb1937a8e7 |
gpui_macos: Migrate display bindings to objc2 (#65714)
Use retained NSScreen objects and objc2 main-thread checks while preserving refresh-rate and display-bounds fallbacks. Release Notes: - [GPUI] Migrated macOS display and display-link bindings to objc2 |
||
|
|
0a9eb81e6c |
gpui_macos: Migrate keyboard and event bindings to objc2 (#65715)
Use typed NSEvent bindings, objc2 Carbon object pointers, and a null-safe NSString conversion. Keep the legacy string conversion and cast the two window event call sites until window bindings migrate. Release Notes: - [GPUI] Migrated macOS keyboard and event bindings to objc2 |
||
|
|
e80af52575 |
Preserve LSP message order (#65856)
Closes https://github.com/zed-industries/zed/issues/65667 Release Notes: - Fixed LSP formatting requests bringing in incorrect changes sometimes |
||
|
|
cc23f35a7d |
Remove deleted files from file finder's history items (#65660)
Closes https://github.com/zed-industries/zed/issues/65657 Release Notes: - Fixed deleted files from file finder's history items not being removed Co-authored-by: Andrew Mitchell <alamottemitchell@gmail.com> |