[mubeda]

What this changes

One batch of fixes for the issues that were open on 2026-11-13, planned in docs/plans/2026-11-13-issue-batch-348-358/ (triage, evidence, spec, plan, phases, progress log, reviews). Every item of every issue was checked against the code first and got one disposition: fixed with a test that failed first, decided by a recorded test or probe, or recorded as a limit in code or docs.

Issue	Change
#348	A terminal re-measures and repaints after a font loads late, and one created after the server's settings arrived repaints once the bundled font is in. A development-only record of size decisions holds no terminal content.
#351	Expanding a reasoning row that the reader scrolled to no longer pulls the chat to its end.
#352	The removal drain test waits on its condition instead of a 2-second bound.
#353	Other AppImage mounts are hidden from Host dashboard volumes; the packaged e2e harness recovers from a dead webview before every test; a re-attach no longer sends a transient terminal size; Fit says when the server kept the size; the Host dashboard's values view no longer redraws on the clock tick; terminals count widths by Unicode 11.
#354	Sign-in guidance appears only for Git's own fatal: line; the pull and hook limits of the index-lock cleanup are recorded; two doc entries added.
#355	Pull-request lookups end at quiesce, a failed lookup waits 5 seconds, a remembered pull request is cleared by a failed lookup on data older than 5 minutes, and a status-neutral local change that outdates a running lookup triggers a new one.
#356	Worktree creation records are deleted by id, both creation-rollback arms reach the profiler audit, and a record under a missing parent directory (an unmounted volume) is kept.
#357	The unsaved-edits block around a remote update install survives a reload, covers a Files source opened during the install, and never ends on a failed state query; the installing dialog has Check again; "Manage devices" is switched off while the install runs.
#358	A popup hides a macOS native preview only where they overlap; a dedicated terminalSizeHandOver capability; a new terminal retries WebGL; refused pairings are logged at warn with a fixed class; the model picker reopens empty; the read-only Markdown view loads relative images from the workspace.
#360	The app no longer falls to "Something went wrong" (React error 185) when a closed branch picker or model picker receives a new list. With the Profiler on, an error the error screen catches is recorded as client.render-error.
#361	Share this host lists this computer's private addresses while remote access is still local-only. On Windows without administrator rights, BiBCode asks through the Windows permission prompt for its firewall rule, or says what is missing and offers Retry. BiBCode no longer manages Tailscale Serve.
#366	Timestamps sent by the web client and those of imported sessions are re-written in the server's nine-digit form before they are stored, so they sort correctly beside server rows. Only the known timestamp fields of each command are touched; answers, tool data and message text stay as sent (follow-up of #362, filed and fixed here).
#362	An expired pairing link, pairing offer or session is refused from the moment it expires, and the OAuth cleanup no longer deletes a refresh token up to a second early. Every timestamp the server stores and compares as text is now written by one function at a fixed width of nine fraction digits (found here, filed and fixed here).

#350 (a server-side terminal screen) was closed as not planned by the owner. Its design, three design reviews and the results of a serializer proof are kept in docs/plans/2026-11-13-terminal-screen-350/.

Diagnoses worth reading

• Windows: expanding a reasoning row scrolls the chat list to its end #351. A trace from the Windows VM ruled out the three paths first suspected. The scroll came from the live-follow of a still-streaming thought, about 150 ms after the toggle: a scroll away from the end that no gesture made never switched "follow the end" off. It is a timing condition, not a Windows one.
• Git Manager: app crashes with React #185 after switching branch and fetching in a large repo #360. Reproduced with the real list library and the real popover: open the branch picker, close it, replace the refs. It fails at 1,400 branches and passes at 1,200. The library recorded a 0 by 0 rectangle of a hidden row as its size; with hidden rows measuring 30 px the same test passes at 4,000. The model picker fails the same way at 50 models. The guard is in both bundle formats of the patched library, and the branch list is rendered only while its picker is open or closing.
• Auth: an expired pairing link or session can be accepted for up to a second after it expires #362. A pairing-refusal test failed once under load. It was not the test: expiry times were stored as ISO text with trailing zeros dropped and compared as text, so a whole-second time sorted after times up to a second later. Before the fix an already-expired pairing was accepted in 21 of 200 runs under load; after it, 0 of 200. The last review found the same form in about ten more writers (the OAuth cleanup, retention cutoffs, the orchestration clock). A first fix at millisecond width made two events in one millisecond tie, which a review caught (the wrong latest assistant message; a skipped check after a compaction), so the final form keeps nanoseconds at a fixed width.
• Follow-ups from the #349 validation rounds #353 (terminal size transient). The test recorded two size claims on re-attach (85x33, then 85x30). A re-measured font, a pixel-ratio change and the bundled-font load now wait for the layout to settle.

Decisions by the owner

• Terminal re-attach: keep a server-side screen state instead of replaying a raw byte tail #350 closed as not planned.
• Git Manager: app crashes with React #185 after switching branch and fetching in a large repo #360's follow-up (the profiler event) and Remote access: Share this host hides LAN/VPN addresses while local-only, and generating a code fails on Windows #361 added to this pull request.
• Remote access: Share this host hides LAN/VPN addresses while local-only, and generating a code fails on Windows #361 on Windows: the Windows permission prompt, over a message only or an installer-created rule.
• Tailscale Serve is not managed from BiBCode; reading Tailscale's addresses stays.

Agent's calls recorded in the spec for the owner to confirm: "Manage devices" is disabled during an install instead of being shown above the installing dialog; a failed lookup's 5-second backoff and the 5-minute rule for a remembered pull request; a creation record is reaped only when its parent directory exists; pull stays out of the index-lock cleanup list; tooltips take no preview lease.

Limits, stated plainly

• Remote access: Share this host hides LAN/VPN addresses while local-only, and generating a code fails on Windows #361, the Windows permission prompt has never been clicked. Detection, the single prompt, a decline, a timeout and the two-prompt replace are proven with a stand-in for the firewall command. The Windows-only code compiles and passes Clippy in a Windows 11 VM, and a probe there sent hostile program paths (spaces, straight and curly quotes, $, backtick, &, %, non-ASCII) through the same ShellExecuteExW call to a stand-in program and got them back byte for byte. How netsh itself reads its arguments, and an approved prompt ending in a pairing code, need a person at a Windows screen.
• Remote access: Share this host hides LAN/VPN addresses while local-only, and generating a code fails on Windows #361, without administrator rights the rule cannot be removed later. Turning sharing off returns the listener to local-only, which is verified every time, and may leave the program-scoped rule in place. The prompt names "Network Command Shell". A prompt left open past its 120-second deadline can still be approved later and add a rule. The network's type is not detected; the dialog notes that a network marked Public may still block other devices.
• Remote access: Share this host hides LAN/VPN addresses while local-only, and generating a code fails on Windows #361, a present rule is reused only when the firewall's rule store shows it enabled, allowing and inbound for this program (a review comment on this pull request; before that fix a disabled rule was reused). The store is read with reg.exe query; if that read fails, the listing's word is taken. The read was probed in the Windows 11 VM and the code passes Clippy and its tests there; no run disabled BiBCode's own rule on a real desktop.
• Git Manager: app crashes with React #185 after switching branch and fetching in a large repo #360 is reproduced in a DOM test environment, not on the packaged macOS app with the owner's repository. Which of the two lists crashed there is not known. There is no packaged spec for Git Manager.
• client.render-error: a reload inside the server's 5-second minimum between batches, or before the connection can send, loses the record. In the packaged build component names are minified, so the route pattern is the readable part.
• Desktop update install: the leave guard after a reload or a lost event #357: this app does not restart a web process that died, so the block's survival is proven across a page reload only.
• Web and terminal: native-preview cover for popups, a size-claim capability, and lab observations #358: the popup lease is proven against a mocked native surface, not on a real macOS window.
• Auth: an expired pairing link or session can be accepted for up to a second after it expires #362: the server still reads timestamp order as event order, so two rows written in the same clock tick can tie, as before this branch. Rows stored before this change keep their old text, so an already stored pairing or session keeps the window it had (under 1 ms for most, up to 1 second for an expiry on a whole second) until it expires. A client timestamp that does not parse is stored as sent, and a command queued before the upgrade keeps its three-digit text. The timestamp change was not run on Windows.
• AI Terminal: rendering breaks intermittently until a panel is resized #348: no test fires a real browser font load or a real WebGL context loss; both are proven with a real xterm object and stubs.
• Follow-ups from the #349 validation rounds #353: the Linux stall and memory probe found nothing with a trace (longest main-thread stall 146 ms over ten route changes; web process 706 to 769 MiB), so nothing was changed for it.

Validation

Where	What	Result
macOS (dev host), final head	vp check --no-fmt, vp check on every changed file, vp run typecheck, vp run check:contracts, vp run check:markdown-links, cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings	exit 0
macOS	web unit suite	788 files, 11,709 passed, 24 skipped
macOS	contracts, shared, client-runtime, desktop package, scripts	937; 844; 1,312; 164; 942 passed
macOS	cargo test -p bibcode-server; cargo test -p bibcode-desktop	7,387 passed, 0 failed; 560 passed (before the last timestamp commits and the merge of main); both exit 0 with no failed test on the final head
macOS	removal drain test alone under CPU load	50 of 50
macOS	pairing expiry, before and after the fix, under CPU load	21 of 200 failed before; 0 of 200 after
Linux (Docker lab, Ubuntu 22.04, WebKitGTK, xvfb), Profiler on	packaged default list	37 passed, 0 failed, 4 skipped (macOS-only cases), on three builds of this branch
Linux	web process killed during a test (four runs)	no hook error; app relaunched once; next test started about 5 s later
Linux	reload during an allowed install	block and dialog restored
Windows 11 arm64 (VM)	removal drain test in 12 whole-suite runs	12 of 12
Windows	packaged chat-reasoning-rows on the fix	3 of 3 (1014 px of movement before the fix)
Windows	build, Clippy, vp check --no-fmt, desktop tests, firewall argument probe (at the commit before the timestamp widening and the merge of main)	exit 0; 471 desktop tests passed; probe passed with no prompt. One whole-suite server run: 6,296 passed, 1 failed (a load test's websocket aborted by Windows, not touched here)
Windows	profiler audit test, 30 runs	30 of 30 (it failed 3 of 12 whole-suite runs before its fixed 1-second wait was replaced)
Released server	current client against v0.15.3 (checksum verified): capabilities, keybindings, attach, resize	legacy path, no error

Two more tests that depended on timing were fixed along the way, without product changes: the profiler worktree audit test (a fixed 1-second wait) and a GitLab poller test (a paused clock moved while a child process was still running; 24 of 160 failed under load before, 0 of 160 after).

Whole-suite runs on Windows at an earlier commit also failed tests this branch does not touch, on a VM sharing the host with other builds: two Git index-lock tests (background_status_does_not_break_a_concurrent_git_add and its linked-worktree variant) and shutdown or single-flight timeouts. On macOS two script tests with a 5-second limit (ci-platform-contract.test.ts, bibcode-identity.test.ts) timed out inside a loaded full run and passed alone. Neither is changed here. On Windows, whole-tree vp check reported formatting in files under docs/plans only, which pass on macOS; CI decides.

Reviews: the plan had two Codex reviews (22 and 13 findings, all handled); the code had Codex reviews of the whole branch and of each later addition, with every finding verified against the code and fixed or rejected with a reason in the plan log.

Profiler: on for every packaged run. No finding caused by this change. Pre-existing on the labs: client.stall and one cpu.sustained on the web process under software rendering, and rpc.slow for Git ref listing and partial staging on the Windows VM. Profiler coverage of the change: a new quiesce step pull-request-lookups (timed by the existing step marks); a new worktree.removal mode create-rollback; a new client event and finding client.render-error through the existing client batch, redacted on both sides; new warn logs carry fixed classes and never a credential, URL or path; the Tailscale status read reports its exit.

React and UI.md: changed components and hooks were reviewed against vercel-react-best-practices, and changed screens, dialogs and messages against UI.md.

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

Closes #348, closes #351, closes #352, closes #353, closes #354, closes #355, closes #356, closes #357, closes #358, closes #360, closes #361, closes #362, closes #366.

🤖 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 2026-11-14T00:19:22.489555Z	1724fac	PR opened
🔒 Security Review	✅ Completed 2026-11-14T00:25:40.238746Z	1724fac	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: 1724face0c

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".