docs(workspace-panels): record the collapsed commit history and its row toggle
- The modal seeded `selectedHash` from the first commit, so History opened with
one commit expanded and a second click could not close it. The page now opens
with every row collapsed, one hash in `useCommitExpansion` is the open row
(`''` = none), and a row toggles itself.
harnesses: recorded the per-platform lease primitive and its measurement
- The lease is `flock` on macOS and an open-file-description lock on Linux
since omp 19.0.1; Linux keeps the two lock spaces apart, so a `flock` probe
reports a live lease as free (measured in a container: `flock` free,
`F_OFD_GETLK` `F_WRLCK`).
- Naming the primitive in the panel page too, since it is the file the stats
plugin walks.
docs(workspace): recorded the repo picker portal and label width
- the list must be portaled because the right panel clips at 268px, and the trigger no longer caps its label at 130px
- measured at a 268px Source Control column in the browser: the 352px list drew outside the panel instead of being clipped
docs(workspace-panels): refresh contract, session-state fallback trap, and the refresh glyph copies
- A panel's own Refresh must re-read every page it draws, and a write must re-read what it changed: git-forge now reloads list + open detail on Refresh and on a write, command-runner also resyncs its run topic, stats also resyncs the topic after its forced sample.
- `useSessionState`'s fallback must be a stable reference: an inline `{}` re-arms the read effect every render and writes a fresh fallback into state — measured at 301 renders and climbing, which hung the Commands panel's first app-level test for 240 s.
- A plugin glyph that mirrors a host icon copies lucide's geometry as separate subpaths: merging the four `d`s of `RefreshCw` changed 82 of 147 456 raster pixels at the arrow tips, while the array form measured 0 (lucide 1.47.0).
docs(workspace-panels): document the synthesized root row and explicit root pick
- Discovery omits '.' for a non-repo root, so the picker synthesizes the row via repoRows().
- resolveRepoForPanel now takes null for "no pick"; an explicit '.' is honored even when the root is not a repo.
docs(terminal): record the copy paths per platform
The macOS shell's Edit menu claims Cmd+C, so the page only sees the native copy: action's DOM copy event; Windows and Linux need Ctrl+Shift+C/V because ghostty-web binds Meta+C only and Ctrl+C must stay SIGINT. Right-click Copy is ghostty's contextmenu handler, which puts the selection in the hidden textarea.
docs(workspace-panels): a new session opens Files, never a stored global pick
Records the measured failure behind the layout fix: `app_settings.activeRightPanel`
still held `browser` from the pre-per-session builds, and seeding a new
session's slot from it drew the Browser Agent instead of the Files tree. Names
`DEFAULT_RIGHT_PANEL_TYPE` as the only source and the e2e proof
(`aria-pressed` false → true).
docs(files): the tar host-spec trap catches test fixtures too
The bz2/xz fixtures packed with an absolute `C:\…` path and failed on the
Windows gate exactly as the product did; the bullet now says a fixture that
shells out to tar uses the temp dir as its cwd and relative names.
docs(workspace): record why the tarball's decompressor is named explicitly
The stdin fix alone was not enough: a stream is not sniffed the way a named file
is, and the Linux gate refused a `.tar.gz` that the same command opened by path.
The bullet now carries that measurement and the per-kind flags.
docs: correct the release re-run claim and record the tar host-spec trap
The stranded-release bullet said a re-run of `Release the app` "cannot fix it"
because semantic-release sees the tag and exits. That is only true when `main`
has not moved: the job fast-forwards to `main`'s tip first, so a re-run with new
commits cuts a release from a commit the run never gated — measured on
2026-11-12, when re-running 43120058573 cut v0.29.0 from a red-gated commit.
The Files panel's archive verbs get their own bullet: GNU tar reads a `C:\…`
argument as a remote host, so the archive now arrives on stdin, and the tool
lookup moved after the request's validation.
docs(panels): record the ACP session lease the Stats panel reads
The panel named an ACP chat from argv, which only `omp acp` matched: the
other 12 harnesses were drawn as a bare `npx` row inside Kabmo's group.
The rule now lives in `session-lease.server.ts` and the plugin reader, and
the measurement (12 of 13) is what a future reader needs to avoid
re-deriving it from the launch table.
docs(workspace-panels): record the terminal viewport hold during a stream
ghostty-web's `writeInternal` ends with `viewportY !== 0 && scrollToBottom()`,
so every PTY chunk re-anchored the terminal panel to the live end: a wheel-up
during `tail -F` was undone by the next frame, and a two-row drag during a
100 lines/s stream copied twelve lines. The page now records the rule that
fixes it, the three states that hold the view, why the hold is an absolute
buffer row, and the re-measured numbers.
docs(workspace-panels): record the terminal focus-on-open rule
The terminal panel now takes focus when it becomes the visible view, and the
request survives the async ghostty-web mount. The grid announcer no longer
focuses on `onReady`, because that stole focus back from whatever the user
clicked in the 120 ms before the grid settled.
docs(terminal): record why the shell starts on open, not on mount
The measured evidence (18 idle shells under the Kabmo pid, bursts of
seven per two minutes, a minted terminal.id per visited session), the
connect/persist decision table, and how `terminal.opened` keeps an
opened shell for the rest of the session.
docs(workspace-panels): record the markdown tokenizer stall
The diff panel's remaining "it takes seconds" case was the tokenizer, not the
diff or the server: Shiki's markdown grammar is superlinear on a long line
full of spans under the JS regex engine (6.3 s per 140-line viewport at the
12 000-char guard, 12-38 ms at 1 000). Records the measurements and the
per-language guard that fixes it.
docs(terminal): match the wiki to the ghostty-web renderer
The terminal page still described xterm: `XTERM_GLYPH_FALLBACKS` (renamed
to `TERMINAL_GLYPH_FALLBACKS`), an atlas refresh that no longer exists
(`refreshAtlasWhenNerdFaceLoads` / `clearTextureAtlas`), `onBinary`, and
a Cmd key that xterm discards. It now states what ships, with the
measurements each rule came from:
- both faces are awaited and a late one re-measures the grid, because the
renderer measures its cell box at `open()`;
- the cell box is Kabmo's `Mg` ink plus 4 px above / 1 px below (16 px),
not the vendor's 13 px that ends on the descender and clips every SGR
background block;
- the terminal paints no surface and the fit drops the addon's 15 px
scrollbar reservation, so the grid uses the whole panel;
- a grid is announced only from a fit that measured the container, is at
least 10x5, and held still for 120 ms — `resize 5x37` and the 80x24
default both reached a shell, whose wrapped prompt redraws then lived
in the scrollback and replayed on every reload;
- PTY frames must leave as binary frames; the JSON byte-map spelling was
the ceiling (4.58 MB/s against 34.8 MB/s).
The neighbouring pages that named `.xterm`, a WebGL canvas, an xterm
buffer and an xterm-owned scrollbar follow, and M21 joins the index.
docs(workspace-panels): record the Files panel reload on reopen
The right-panel stack keeps hidden views mounted, so 'active' had to mean selected AND the column open for FileExplorer to notice it was shown again. Names the regression spec and the harness helper that reproduce it.
docs(workspace-panels): record why the Sync button rotates its own glyph
The in-flight mark is a rotation of `RefreshCw`, not a swapped-in spinner —
a different icon changes the control's identity mid-press — and it is derived
from the running action rather than the panel's `isBusy`, which would also
spin the badge for a commit, a stage or a status read. The e2e spec holds the
POST open to observe the state, because an unheld pull+push finished before a
static glyph could be told apart from an unpressed button.
docs(diff-panel): a small file can be slow when the change is dense
The reporter's slow case was under 1 000 lines, and line count is the wrong
axis: the badge's own scan cost 667 ms at 200 lines rewritten whole, 2.9 s at
400 and 7.0 s at 600, against 1 ms for 900 lines with four edits. An idle
harness opens the same diff in 276 ms, so the mount was never what a reader
noticed.
docs(diff-panel): record what a load costs and the badge's second diff
The badge re-diffed the pair with no `diffConfig` while the view built its
chunks with `scanLimit: 500`, so the panel paid a second, far more expensive
diff and could print a number that disagreed with the rows: 4.0 s and 540/540
against the view's 26 951 on a 27 000-line pair, and 55 s on a 2 000-line pair
that was wholly rewritten.
Also records the memoized whitespace fold, the remaining cost (the view's own
diff, 6.6 s at 200 000 lines, with no size guard on the path) and the
measurement that the server is not the bottleneck (40–145 ms per file diff).
docs(browser): record why the agent panel seeds its first frame
Measured on Chromium 1248: a static tab delivers 0 screencast frames at any
everyNthFrame, while captureScreenshot answers in 8-40ms with the page's real
pixels. Also records the SSE backpressure drop that re-created the hang
downstream of a correct viewer.
docs(workspace-panels): record the collapsed-chain rules and the drop highlight
A collapsed chain is one row standing for several directories, and everything
that went wrong in the Files panel followed from treating it as one path:
a drop aimed at a level inside the label landed in the deepest directory, the
Rename field could not address a level at all, and the drag highlight was drawn
from "would a drop be legal here" so the whole tree painted at once.
Records the five rules and the two measurements behind them: unfold for the
navigation cost of the fold, `isChainRow` as the single "does this row hide
levels?" predicate (a label comparison, not `name !== path`), the open/close
pairing of Unfold and Fold, the three rename cases that preserve or split a
chain, and the harness rule that a rename must move a subtree.
docs(panels): record the exists-probe scope rule
The "file is gone" probe asks the same `root` + `repo` + `path` the read route
does, and the phone dropped `repo` — measured against a live server, a
nested-repo file answered `exists:false` for a path that had just been read.
Which tabs may be probed at all is one predicate in the hook, because the two
layouts had applied it differently.
docs(panels): record that a plugin's floating async loop can kill the server
A background service's rejection is not always catchable at the source: the
shipped stats plugin starts a floating `async` IIFE inside `start()`, so the
promise the loader would attach a `.catch` to is `undefined`. The entry names
that shape, the measured `SQLITE_IOERR_VNODE` it produced when a data directory
was deleted under a running server, and the process-boundary guard that now logs
it instead of exiting 1 — plus why `uncaughtException` stays unhandled.
docs(panels): the scrollbar arm is bound to the container the input landed in
A wheel that cannot move its container fires no scroll, so its unspent arm was
spendable by a programmatic scroll anywhere. Records the measurement and the
binding rule.
docs(panels): record the scrollbar gate as an arm, not a time window
The rule changed twice for the same symptom — the bar staying dark while the
reader scrolled — so the page now carries both failures and the measurement
that separates them: `Date.now()` inside the handler (Chrome), then
`Event.timeStamp` on both sides (still WebKit, which stamps a scroll at
dispatch: 703ms for a 700ms blocked frame). An arm asks no clock question.
docs(memory): record the byte budget and the canvas frame sink
Both entries carry the measurement that found the defect, so the next
reader can re-derive the numbers instead of trusting a constant.
The editor's cache was bounded by tab count alone, which is not a memory
bound: a text file's ceiling is 32 MiB, so twelve of them is up to 384 MB
of document text plus each tab's undo history. It now also holds an 8 MiB
budget, and the document being opened is never the one evicted.
The screencast handed every frame to an `<img>`; WebKit never releases the
decoded image behind a replaced `src` — measured at ~76 KB of WebContent
footprint per frame (148 MB for 1 760 frames) against 28 MB for the same
frames decoded to an ImageBitmap, drawn and closed.
docs: split the manual into a wiki