docs(build-test): record the fake omp's agentInvoked ack contract
- A builtin that opens no turn must be acked `agentInvoked:false`, and that ack
must be the only response. The wrapper reads `true` as "this dispatch claimed
the turn slot" and holds `promptRunning` until the agent-start watchdog fires,
while the `prompt_result {agentInvoked:false}` that follows is suppressed by
the run-attribution rule — so nothing clears the flag early.
- Measured on release run 44249018348: the fake acked every prompt `true`, the
client's `/kabmo-mode` republish left the session reporting busy for ~10 s
after the answer landed, and `rename.e2e.ts` was refused with "This session is
working" on a session that had finished — 4 of 10 shards red, 1 flake in 6
locally, 0 in 10 after the fix.
docs(build-test): record the mounted-tree leak behind the Windows shard failure
- Release run 44213261035 failed 8 cases in `list-slots.test.ts` with `ReferenceError: document is not defined`; the leak was `MessageList.test.ts` leaving live trees subscribed to the list-slot store.
- Names the shape to grep for (a file that calls `render()` with no `render(null, …)`) and the rule that an order-dependent failure names the file that surfaced it, not the file that leaked.
docs(build-test-release): record the three-way Linux shard split
- The browser suite now runs 3 shards on Linux (60/53/50 tests) beside 4 on Windows, so the page states the new counts and that Windows, not Linux, is the bound of the run.
docs: record the reparenting race a spawned-holder helper has to wait out
The peer-relay case waited for the lease and returned, then wrote an instance
record naming a guessed pid, while the guard matches the holder's PARENT — a
moment later. Measured: one Linux gate answering locally with no owner call.
docs: record the editor.run ownership rule and two test-global leaks
App-Shell-and-Layout: a command id that throws on a duplicate must be
registered by the runtime when the surface can be mounted twice — the split
editor's second pane threw `command "editor.run" is already registered`, and
the dev overlay then blocked the next click (measured screenshot: 1280x720).
Build-Test-and-Release: `fetch` joins the globals a suite must restore, with
the pair that reproduces the CLI suite's 5s timeouts, and the persisted-slice
ids a harness mints must be process-wide rather than per-file.
docs: record the split build, the generated stylesheet asset and the DB write path
Build-Test-and-Release: two builds and why one cannot split the client without
publishing server chunks under `static/`; the shell is a Bun HTML route in
development and `dist/client/index.html` in production; the measured bundle
figures.
App-Shell-and-Layout: the two generated stylesheets are a linked,
content-hashed asset, the layer-order reason the link comes first, and the
`link[data-theme-css]` repair guard.
Runtime-and-Toolchain: WAL with `synchronous=NORMAL` and `writeMany`'s chunked
transactions, with the measured per-write costs.
docs: record the DOM-globals pair, the voided in-flight read, and platform-named tests
The Release gate of 2026-11-13 failed 20 cases on two legs, none of them in
the file that caused it. Three rules keep the next reader from re-breaking it:
a suite restores the window it installs through `pristine-globals` and that
list carries `EventTarget` beside `Event`; a module-level client cache is
voided together with its in-flight read; and a test that takes a platform
argument builds its expectation with that platform's rules.
docs(e2e): record the terminal stream spec's measured Windows limit
The platform-limit bullet warns that a skip must be reproduced against the real
spawn and must name the exact cause, so the one skip that survived that test now
carries its measurement: `cmd.exe` turns the POSIX stream loop into arguments to
Git-for-Windows' `clear.exe` (screenshot in run 43112230252's artifact), and
ghostty-web binds copy to `Meta+C` only.
It also records the product gap the skip exposes: a Windows desktop has no
keyboard copy in the terminal, because `TerminalKeyBar` renders only on touch.
docs: record why the browser suite is sharded and why `bun test --shard` is not
The release gate was 87% of a 33-minute run, all of it the Windows browser
leg. Sharding is the fix and `workers` is not: measured, `--workers=2` bought
12% and made two specs flake, because two servers and two browsers contend for
one CPU. A shard is a separate machine, so it keeps the one-worker-per-machine
property the config depends on.
Also records the blocker found while sharding the unit gate: `bun test
--shard` splits by file, and this suite is not shard-safe — 9 failures across
the two shards against 0 in the unsharded run, caused by cross-file env
leakage that makes `getDb()` seed workspace folders into a database the next
file expects to be empty.
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: record the one-workflow-per-push and staged shell bump rules
Two pipeline facts changed and each came with a measurement, so both belong here
rather than only in the workflow comments a future reader may not open.
A push to `main` now raises exactly one workflow. `ci.yml` and `e2e.yml` carried
`push: branches: [main]` beside `release.yml` and re-ran the same reusable
workflows on the same commit, and the duplicate pair had already disagreed: on
26333b1 the CI leg's `Gates / Checks (Windows)` failed on an `EBUSY` temp rmdir
while the Release leg of the identical job passed.
A shell version bump reaches `main` only after every shell build passed, on the
staging branch `release/shell-bump`; `shell-commit` fast-forwards it once the APK,
IPA, Windows, macOS and Linux builds are green. Committing first was the bug: a
failed build left a bumped version on `main` for an artifact that never existed.
The e2e page's claim that the browser layer runs on every push is corrected to
what the pipeline now does: label-gated pull requests, nightly, and as a gate of
the release run.
docs(build): record how the compiled-fixture cache and unit temp dirs leaked
Two directories under `$TMPDIR` grew without bound and neither showed up as a
failure. The compiled-fixture cache is keyed by a hash of the stub's source, so a
baked per-run path missed it every run: ~200 MB per `bun test`, 40 GB on one
machine. Unit suites left 30 entries per run, and `process.on('exit', …)` — the
hook one of them relied on — is never fired by `bun test`.
Both entries carry the measurement that found them and the audit command that
reproduces it (`TMPDIR=/tmp/audit bun test && ls /tmp/audit`), because the leak
is invisible against a developer's already-full temp directory.
docs: split the manual into a wiki