Repository navigation
Linux 原生伴生架构 + 多屏截屏 + 阅读线索上下文 - #1
Merged
Merged
Conversation
- add root Cargo workspace and LF normalization - decouple glint-core from Tauri via StreamSink and path-based config - fix Tauri 2.12 tray icon API - separate contract vs runtime acceptance; skip unimplemented Slint/placeholder tests - fix single-newline markdown rendering Co-Authored-By: Claude Code <noreply@anthropic.com>
- add session detection, PRIMARY selection sensor (wl-clipboard-rs + x11rb), and KGlobalAccel/portal hotkey engines - add glint-cli with session/capture/explain/hotkeys commands - replace canned offline answers with an explicit not-configured state - document measured runtime evidence and the .desktop component-id pitfall Co-Authored-By: Claude Code <noreply@anthropic.com>
…ming - replace placeholder screenshot with xdg-desktop-portal capture - infer widgets from JSON shape when the model omits the widget: tag - fix WebKitGTK flex min-width collapse in the answer area - add pin button and shell-side pin state - stream the answer without a blocking tool round; tools opt-in - share one config path between CLI and shell Co-Authored-By: Claude Code <noreply@anthropic.com>
- add per-code-block copy with delegated clicks and stream-safe state - infer widget type from JSON shape when the model omits the tag - verify compositor placement and log when it is ignored - add scripts/glint-x11.sh for cursor-adjacent placement and real AOT - fix attr encoding so copied code round-trips exactly Co-Authored-By: Claude Code <noreply@anthropic.com>
- re-apply the card size after the surface is mapped (hidden windows report 0x0) - read the position back on a delayed task instead of before mapping - raise the default card to 440x560 so an answer is readable without scrolling - revert the body inset; keep the original edge-to-edge card Co-Authored-By: Claude Code <noreply@anthropic.com>
The screenshot portal returns the whole virtual desktop, but KWin confines a toplevel overlay to one monitor's work area (measured: a 2987x3280 request came back 2987x1610 at y=70). Sizing the overlay to the virtual desktop and drawing the image into it squeezed both displays into one frame. - Overlay now covers only the monitor under the pointer. - The capture image is cropped 1:1 to that monitor slice; nothing is scaled. - Rust reports the mapped physical window rect after show, so the crop does not depend on `window.screenX` (CSS pixels, wrong under fractional scaling). - Placement follows the monitor under the cursor instead of the HUD's stale `current_monitor()`. - The HUD is hidden before the portal grabs the desktop and restored on cancel/error, so the card no longer lands in its own screenshot. - A screenshot appends an image turn to the live conversation instead of resetting it; only an empty card starts a new session. - Geometry helpers (bounding_box, monitor_at_point) live in `placement` with tests for multi-monitor, negative origins and gaps. - `glint-cli screen [out.png]` verifies portal dimensions and dumps the frame. Co-Authored-By: Claude Code <noreply@anthropic.com>
Sessions used to reset on every trigger, so a screenshot could never continue the discussion it was taken for. Making them always continue would silently merge unrelated pages, which users cannot see and which degrades answers. Neither is acceptable for a reading companion. - `glint_core::host::HostIdentity`: stable app class (KWin `resourceClass`, `xdotool` fallback), RAM-only title, `doc_key` digest of the normalised title. Generic titles produce no key. - `glint_core::context::plan_turn`: continue only within the same app and document, within 30 min (5 min for screenshots), under 8 turns. Every edge case fails toward a new thread. - Continued threads get a visible divider and per-turn labels; continuity is never silent. - The host label is frozen per thread, shortened, and enters the prompt as an escaped `<host_label>` data block with a "not an instruction" rule. Titles are attacker-controlled and no longer sit in the instruction slot. - Selections are wrapped in `<selection>` with bidi/zero-width stripping. - Host identity is read before `show()`; otherwise KWin reports Glint as the foreground app and every selection looks like the same document. `is_glint_window` also refuses to record Glint as a host. - History stores `app_class` + `doc_key` instead of the raw window title. - ADR 0013 documents the model, the rule and the rejected alternatives. Co-Authored-By: Claude Code <noreply@anthropic.com>
…e query `org.kde.KWin.queryWindowInfo` is the D-Bus endpoint behind "click a window to identify it": calling it puts the compositor into selection mode, turns the cursor into a crosshair and waits for a click. Every Alt+Q therefore hijacked the pointer, and an unanswered query could leave the compositor in that mode. Read `_NET_ACTIVE_WINDOW` and then `WM_CLASS` / `_NET_WM_NAME` with `xprop` instead — same data, no input capture, no modal state. Also fixes the caption parsing that motivated the JSON experiment: the plain `busctl` dump splits a title on whitespace, so a multi-word caption was truncated to its first word and `doc_key` hashed a fragment. xprop returns the whole quoted string. Co-Authored-By: Claude Code <noreply@anthropic.com>
The repo had no CI, no git hook and no single check entry point, so breakage was only found by whoever happened to build next. Four real defects this session were found by hand: an infinite recursion, a host read after window focus, an interactive KWin call, and Windows code that has never been compiled. - `.github/workflows/ci.yml`: frontend typecheck/unit/contract/build; `glint-core` tests on Linux/Windows/macOS; Tauri shell build, tests and `clippy -D warnings` on all three. - `scripts/check.sh`: the same sequence locally, so a push fails here first. `scripts/install-hooks.sh` wires it as a pre-push hook. - `pnpm check` runs the frontend half. - Rust is now `cargo fmt`ed and clippy-clean under `-D warnings` (28 warnings fixed: collapsible ifs, manual checked division, a complex type, a dead test variable). - Windows/macOS gain real `HostIdentity` implementations (process executable name / bundle identifier) instead of an empty app class, so thread continuity works off Linux too. The frontend falls back to a shorter time window when a platform reports no class. Co-Authored-By: Claude Code <noreply@anthropic.com>
…lper Retro follow-ups from this session's four hand-found defects. - AGENTS.md (with CLAUDE.md importing it): pointers only — check command, layout, the `default-members` trap, ADR index, and "ask before launching the GUI" because starting the app repeatedly disrupted the user. - CODING_STANDARDS.md: four judgement calls a linter cannot enforce — OS integration must be non-interactive, prompt tests assert markers not prose, a failing test asks which side is wrong, platform code is only real once CI compiles it. - scripts/shot.sh: captures and crops by physical desktop coordinates, auto-detecting the capture scale. A cropped HUD shot is 22 KB instead of a 2.7 MB full-desktop image, which is what made UI verification expensive. - tests/README.md points at `./scripts/check.sh` and names the command that actually runs the shell tests. Co-Authored-By: Claude Code <noreply@anthropic.com>
The initial run exposed exactly what CI was added to catch. - Windows: `HWND` is `*mut c_void`, not `isize`. Use the real type and `is_null()` instead of casting through `as _`. - macOS/Linux core tests: two tests used Linux-only modules (`linux::KGlobalAccelEngine`, `percent_decode`) without a cfg gate, so `cargo test -p glint-core` failed to compile off Linux. Gate the tests and the import, not the assertions. - Unused-import warnings on Windows for `is_kde` / `detect_session`. - Contract suite: three tests shell out to `cargo test` inside src-tauri. The frontend job has no GTK/WebKit headers, so they failed for an environment reason. They now skip with a reason when the shell cannot be built, driven by `GLINT_SHELL_BUILDABLE` (auto-detected via pkg-config). The tauri job still compiles and tests the shell, so coverage is unchanged. - Ubuntu dependency install: `libappindicator3-dev` and `libayatana-appindicator3-dev` conflict; install only the latter. Verified locally for the Rust halves: `cargo check -p glint-core` for x86_64-linux, x86_64-pc-windows-msvc and aarch64-apple-darwin. Co-Authored-By: Claude Code <noreply@anthropic.com>
…tests - `default_config_path` test compared a string suffix; Windows uses `\`. Compare path components instead. - The Windows clipboard test asserted `seq > 0`, but a CI runner has no interactive clipboard, so 0 is legitimate. Assert the API resolves. - macOS clippy: `Command` import in linux.rs, `Arc` in lib.rs and the `app_handle` parameter in macos.rs are all platform-gated. Gate the imports, rename the parameter. Linux and Windows core are checked locally; the macOS shell can only be compiled on a macOS runner (objc2 needs Apple's C toolchain). Co-Authored-By: Claude Code <noreply@anthropic.com>
`clippy -D warnings` on macOS flagged symbols that only exist for the Linux path: GLINT_APP_ID, virtual_desktop_bounds, active_monitor_rect, and the xprop parsers in platform/linux.rs. Gate them with `#[cfg(target_os = "linux")]`; `placement::bounding_box` is pure geometry kept for tests, so it gets a targeted allow instead of a cfg that would remove it from the test build. Co-Authored-By: Claude Code <noreply@anthropic.com>
`default_config_path` only knew `XDG_CONFIG_HOME` and `HOME`. Windows has neither, so it fell through to `glint_config.json` in the working directory — the CLI and the GUI would read different files, which is the exact split this function exists to prevent (and it already caused one "why does the GUI work but the CLI says no API key" bug). Resolve per platform the way `app_config_dir()` does: XDG on Linux, `%APPDATA%` on Windows, `~/Library/Application Support` on macOS. The test caught it because it asserts the path shape, not a string suffix. Co-Authored-By: Claude Code <noreply@anthropic.com>
- `clippy -p glint-core` for the Windows target found four real issues invisible on Linux: two `as u32` casts of an already-u32 constant, and a manual BGRA→RGBA loop that `as_chunks_mut` expresses directly. - The same for macOS found a collapsible `if` in both copies of the macOS cursor helper. - `scripts/check.sh` now runs the core lints for both cross targets when the targets are installed, so platform-only code is checked before a push rather than one platform per CI round trip. Co-Authored-By: Claude Code <noreply@anthropic.com>
`clippy -D warnings` on Windows and macOS flagged the same function five times for nested `if let`. Extract `clipboard_text()` and return the candidate from each branch instead of nesting; behaviour is unchanged (sequence number when the platform has one, text comparison otherwise). Also gate the placeholder-screenshot import to the platforms that use it and drop three redundant `as u32` casts in the Win32 hotkey registration. Co-Authored-By: Claude Code <noreply@anthropic.com>
Same nested `if let` shape that was fixed in the Linux path; the Windows/macOS build flagged it. Both platforms now compile the same expression form. Co-Authored-By: Claude Code <noreply@anthropic.com>
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configuration
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Lines 646 to 650 in 1ae319c
Preserve the captured host anchor through snip submission
The backend captures and emits app_class and doc_key with the snip frame (src-tauri/src/lib.rs), but this submission forwards only the title. Consequently the HUD receives both anchor fields as null, and startNewMultimodalQuery falls back to currentThread's app/document identity. If a user has an existing thread for app/document A, then activates B and snips it within five minutes, B's image is silently appended to A and A's history is sent with it instead of opening a new thread.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…adlines - Snip submission forwards the frame's `app_class`/`doc_key` instead of only the title. The HUD used to fall back to the open thread's identity, so a snip taken from app B within five minutes was appended to app A's history and A's transcript was sent with it. - Continuation now refuses when either side has no app anchor. The old time-window fallback treated "no evidence" as evidence of continuity; a timer cannot distinguish two unrelated documents. - Turns are serialized. A trigger arriving before the previous `Done` used to start a second request against the same shared transcript, producing `U1, U2, A1, A2`; the newest trigger is queued and started only after the current answer is recorded. - Closing the snip overlay through the window manager (Alt+F4) restores the HUD, matching the Esc path. Previously the card stayed hidden. - The PRIMARY read deadline now covers the whole read: poll-and-read in bounded chunks instead of an unbounded `read_to_end` after one poll, so a stalled selection owner cannot hang the hotkey path. Co-Authored-By: Claude Code <noreply@anthropic.com>
Em dashes are the one punctuation habit that makes the comments read as generated. Replace them with periods, commas or colons across the files added in this branch, and rewrite the two judgement rules in CODING_STANDARDS.md as direct sentences instead of bolded labels. Co-Authored-By: Claude Code <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
概述
让 Glint 在 Fedora 45 / KDE Plasma 6.7 / Wayland 上可用，并修掉两个使用中暴露的缺陷。
Linux 原生能力
com.glint.Glint，因为 kglobalacceld 会把以.desktop结尾的名字当作桌面文件动作。wl-clipboard-rs/x11-clipboard），不模拟 Ctrl+C，不写剪贴板。xdg-desktop-portal截屏，编码为 data URL。多屏截屏
截屏门户返回整个虚拟桌面，KWin 把 overlay 限制在单个屏幕的工作区。实测请求 2987×3280，返回 2987×1610，位置 y=70。原实现把整张图压进一个屏幕。
阅读线索
会话原来每次触发都重置。现在按应用和文档锚定：
<host_label>数据块。历史只存应用标识和标题摘要。工程护栏
./scripts/check.sh在本地跑同一序列，包含跨目标 clippy。AGENTS.md和CODING_STANDARDS.md记录导航和判断类规则。scripts/shot.sh按物理坐标裁剪截图。修复的缺陷
queryWindowInfo是交互式接口，调用后光标变成十字并等待点击。改用xprop读_NET_ACTIVE_WINDOW。default_config_path()在 Windows 上读不到HOME，CLI 和 GUI 使用不同配置。HWND类型误用，macOS 和 Linux 的测试缺少 cfg 门禁。验证
./scripts/check.sh通过：fmt、clippy（-D warnings）、core 48 测试、shell 22 测试、前端单测、契约套件 77 通过、生产构建。未完成
文档
🤖 Generated with Claude Code