[mubeda]

What this changes

One batch of fixes and two features, planned in docs/plans/2026-11-09-issue-batch-340-345-290/ (spec, plan, phases, progress log).

Issue	Change
#340	Pipeline views keep their header and selected job readable.
#341	A Git index lock left by a stopped command no longer blocks later Git work; long Git writes can be stopped.
#342	Checking a Markdown task box is one undo step; closing the window or quitting with unsaved Files edits asks first (Save all, Discard, Cancel). Mermaid diagrams render again in rendered Markdown.
#343	A refused push to a remote explains what to do.
#344	Worktree orphan warnings rest on durable creation records, so a worktree BiBCode created is not reported as an orphan.
#345	Rich Markdown editing in Files; Git status keeps arriving during network Git operations.
#290	Custom themes, and five levels in the left panel.
#292	A Host dashboard per environment: CPU, memory and volumes, live and since the server started; a host_metrics MCP tool; volumes the filter hides are listed with the reason and never read.
#293	A Review changes button in Source Control, with review prompts per project, one-off instructions, a shortcut, and terminal reviews on Windows.
#346, #347	The composer's model picker no longer shows stale values, and a chat bound to a provider that is unavailable says so and refuses Send.
#348 (not closed here, see below)	Terminals shown in two windows, a view left above the bottom, and missing characters after returning to a Claude terminal.

Also: the Linux AppImage bundles its GStreamer plugins (audio in Files no longer kills the web process), the PDF preview loads (frame-src now allows blob:), and several tests that raced their own timeouts are fixed.

#348 in detail

Three separate causes were found on a real pair (the Linux desktop app as server, a second client):

1. A window that does not size the terminal was cut off at the bottom. The window used last sizes the terminal; other windows show the whole grid scaled to fit (never below half size) with a notice. The server now answers a resize with what it holds, hands the size over when the sizing window leaves, and names one window to answer terminal queries while nobody owns the size.
2. The view was left above the bottom. xterm 6 stops drawing a terminal whose area is 0 px wide (the right panel squeezing the center column). A row change from the other window then moved the scroll position, and the panel took that for the person scrolling up. The panel now records whether the view followed the output when drawing stops and restores it when drawing resumes. Automatic size requests wait until the terminal's area holds still.
3. Characters were missing after returning to a Claude terminal. Every return rebuilds the screen from the server's raw byte history (1 MiB or 5,000 lines). The Claude CLI's fullscreen renderer writes only changed cells, so once its last full paint has left the history, cells, the screen mode and the input modes are lost. After such a lossy attach to a Claude terminal that BiBCode started, the server now sends SIGCONT to that terminal's own process group, and the CLI re-enters its screen and repaints.

Limits, stated plainly:

• The repaint after a lossy attach is best-effort recovery for Claude terminals that BiBCode starts, on Linux and macOS servers. It is not terminal-state reconstruction. Other programs, Codex terminals and Windows servers behave as before. The lasting fix (the server keeps an emulated screen) is a follow-up.
• A Claude session suspended in its own terminal resumes when a window re-attaches after a lossy history.
• Pressing the scroll bar of a scaled-down terminal takes the size for that window; the mouse wheel does not.
• The pair's client was headless Chromium and the Linux app's WebKitGTK window, not the macOS app. AI Terminal: rendering breaks intermittently until a panel is resized #348 stays open until it is confirmed there.

After the first review

• Index lock ownership ([Bug]: Background Git polling makes the user's own git add/commit fail with "index.lock: File exists" #341). On Linux and macOS, a leftover index.lock is removed after a forced stop only if the killed Git process still held the lock open just before the kill, so a lock another Git process took meanwhile is kept even when it stays unchanged. Windows cannot check this: there a foreign lock that appeared while BiBCode's command held none and stayed unchanged for 2 seconds can still be removed. The Linux check has not run yet outside CI.
• Images in rendered Markdown (Git: refreshing status hangs while a stacked action holds the read gate #345). A relative image loads from the workspace through the same asset and blob path as the Files preview; the saved file keeps the relative path; paths that leave the workspace are refused. Checked with the real editor in unit tests and a new packaged assertion that has not been run; not seen in a real browser. The read-only rendered view (shown while the editor loads, for files over 128 KiB, and for files the editor cannot change safely) still does not load relative images, as on main.

macOS app as client (run after CI went green)

The packaged macOS app (WKWebView) built from this branch was the client of the Linux desktop app running in the Docker lab (Ubuntu 22.04, WebKitGTK, Xvfb, no window manager), Profiler on in both. The server was a container, not a real desktop, and input on the Mac was synthetic.

Check	Result
A window that does not size the terminal (DOM and WebGL renderer)	0 px cut off in 190 checks; scaled with the notice, down to the 0.5 floor
Right panel collapses the terminal to 0 px while the other window changes its rows	at the bottom at once 15 of 15
Click, Shift-click, Fit, then typing on a scaled terminal	size taken; the key arrived 20 of 20
Lossy re-attach of a Claude provider terminal (stand-in)	one SIGCONT per lossy attach 30 of 30, none on complete histories; 27 of 30 screens identical, 3 differ only by live output read a moment apart
The Linux window reloads	size handed to the Mac 3 of 5; in 2 of 5 the Linux window claimed it again at its own reload
Profiler	no finding caused by this change

Found and not fixed here: when the PTY size changes while a window is away and that window returns to a complete history, the raw replay at the new size can leave rows that an attached window does not show (4 stale rows, 5 of 5 runs with the stand-in). A headless xterm.js probe shows the cause is the replay of bytes written for another size; a program that erases the whole screen when it is resized is not affected. Whether the real Claude CLI is affected is not known. The server-side emulated screen (follow-up) removes this class.

Not valid on this pair: whether an idle window takes the size back (every driver call counts as a user gesture, and the lab's window always reports focus). Not run: jumping between two terminals of two worktrees (the second worktree had no terminal).

Compatibility

• Every changed wire shape is an optional field or sits behind a capability, and absent means the old behaviour. This is covered by contract tests and test doubles for both directions; the only mixed-version run on real builds was this client against an older server for the Host dashboard (it shows the "update the server" state). Terminal resize, attach and keybindings were not run against a real older server.
• One known limit: a keybinding for Review changes (sourceControl.review) saved on a server makes a v0.15.3 client unable to load that server's keybindings, as chat.find already does for older clients. The keybindings guide says so.

Validation

The suites below ran on the tree of the last product commit (fix(terminal): ask a Claude terminal to repaint after a lossy attach). The commits after it change three test files and the plan log only; the changed tests were rerun on their own. Commit ids in the plan log predate a history rewrite that removed a local folder path from two early plan commits; the final tree is unchanged by it.

Where	What	Result
macOS (dev host)	cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, vp check, vp run typecheck, vp run check:contracts, vp run check:markdown-links	exit 0
macOS	web unit suite	11,529 passed
macOS	scripts suite; desktop e2e support tests	939 passed; 145 passed
macOS	cargo test -p bibcode-server ; cargo test -p bibcode-desktop	7,340 passed, 1 failed (a stale CSP assertion, fixed in the next commit and rerun: 23 of 23 in its file); 532 passed
Linux (Docker lab, Ubuntu 22.04, WebKitGTK, xvfb)	packaged default list; Host dashboard, Source Control review, PDF preview and footer specs	36 passed; all passed (one footer run timed out under load and passed on rerun; its time allowance was raised)
Linux	cargo test -p bibcode-server terminal::, host_metrics; Clippy	408 passed; 66 passed; exit 0
Windows 11 arm64 (VM)	fmt, Clippy, vp check, typecheck	exit 0
Windows	cargo test -p bibcode-desktop; cargo test -p bibcode-server	443 passed; 6,259 passed, 2 failed (the same stale CSP assertion; and removal_refuses_when_an_admission_does_not_drain_before_the_deadline, which fails about 1 run in 3 on main too)
Windows	the ten review prompt cases, ConPTY resize, mapped network drives in the Host dashboard	pass
Real pair	see the table below

Real pair (Linux desktop app as server, headless Chromium clients, Profiler on):

Check	Before the fixes	Final build
Recorded divider-drag sequence, lost footer	18 of 19 replays	0 of 20
Right panel collapses the terminal to 0 px while the other window changes its rows	view left above the bottom 15 of 15	at the bottom at once 15 of 15, also with no output after the panel closes
A view scrolled up on purpose (wheel, scroll-bar drag)	n/a	kept across the collapse; a wheel scroll after it is not pulled back
A program's scroll region with a fixed footer	broken 59 of 61 checks	intact 61 of 61
Lossy re-attach to a Claude provider terminal (fullscreen, cell-diff stand-in that handles SIGCONT)	0 of 5 correct	5 of 5: 0 differing rows, alternate screen and input modes restored; one signal per lossy attach; PTY size unchanged
Lossy re-attach of a plain shell with a half-typed command; a complete short history; a Codex terminal	n/a	no signal sent in any of them; the shell's line is intact
A window that does not size the terminal	cut off by 85 to 257 px	0 px in every check, scaled with the notice
Click, Shift-click and Fit on a scaled terminal	size taken without focus 15 of 15	size taken and the next key arrives
A focused but idle window (tab shown again, reload; 10 each)	took the size on older builds	0 of 20

Not run: the owner's own Linux desktop as server with the macOS client; the packaged macOS suite (the host's console was locked, which hides the webview); the real Claude CLI on the pair (not signed in inside the test fixture; the owner confirmed by hand on their own server that kill -CONT repairs a terminal with missing characters); x64 Windows and Linux (CI); a focused WebKit window for the idle take-back check.

Profiler: on for every packaged and real-pair run. No finding caused by this change. Pre-existing: main-thread stalls up to 2.7 s and a few slow calls under software rendering, more rpc.slow occurrences for Git ref listing on the Windows VM.

Profiler coverage of the change: the unary server.getHostMetrics and the host metrics stream are registered through the session wrappers and are not long-running; the host collector stops in quiesce before the profiler's teardown and its disk threads are never joined; the resume signal adds no process, RPC or event and logs through tracing; no terminal content or mount point is logged.

React and UI.md review: the changed terminal panel, composer banner and Host dashboard components were reviewed against vercel-react-best-practices and UI.md (five small findings; three fixed: the unavailable-agent banner's guidance, the size notice as a status region, a 100% label for a scaled terminal; two left: the Fit button gives no feedback when no size request can be sent, and the Host dashboard sheet re-renders once a second while open).

Runbooks: docs/testing/cross-platform-validation.md, linux-desktop.md and windows-desktop.md were updated; the others were reviewed and remain accurate.

Known limits and follow-ups (ticketed separately)

• A server-side emulated terminal screen for snapshots (the lasting fix for lossy re-attach; spike results and review requirements recorded).
• On Windows, expanding a reasoning row scrolls the chat list to its end (chat-reasoning-rows, also on main).
• removal_refuses_when_an_admission_does_not_drain_before_the_deadline is timing-sensitive under load (also on main).
• A second BiBCode AppImage mount (another instance, or a stale mount) shows as a read-only volume in the Host dashboard.
• After a web process crash the packaged e2e harness relaunches only once per worker.

Refs #348.
Closes #340, closes #341, closes #342, closes #343, closes #344, closes #345, closes #290, closes #292, closes #293, closes #346, closes #347.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-12T22:33:32.240054Z	9c57548	PR opened
🔒 Security Review	✅ Completed 2026-11-12T22:35:49.506371Z	9c57548	PR opened

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" or "@codex security review".

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]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9c57548ffa

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".