runtime.ts: the runtime is pinned in a compiled constant (version, URL, size,
sha256). The archive is streamed to a .partial file while it is hashed,
continued with a Range request, retried at most 3 times, and the hash is
compared before anything is unpacked; a mismatch deletes the file. It is
unpacked with /usr/bin/tar into a temporary folder and renamed.
The supervisor runs the runtime as a non-detached child with an environment
built from an allowlist by the pure runtimeEnv(), on a port chosen once by
binding. It restarts the child at most 3 times, stops it with SIGTERM, adopts
or replaces its own orphan on the port and answers port-in-use for any other
listener. parseInferenceComputeTotal() reads the device total the runtime
logs when it starts. Under vitest nothing spawns or fetches.
ollama-client.ts: the pull is parsed incrementally by a pure parser (a 23 GB
pull is about 1.4 million lines), with up to 5 retries on an error line or a
cut stream. verifyModel() loads the model with one tiny chat, checks what the
runtime reports (all on the GPU, the chosen context) and makes one tool
probe; each failure is a code.
developer-tools.ts: only /usr/bin/xcode-select -p, bounded, darwin only.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The catalogue rows carry the figures measured on a Mac (the MLX build of
qwen3.6:35b-a3b is 23,105,798,956 bytes, about 37 KB held per context token,
about 3 GB of runner overhead, about 5 GB more for a moment while a long
prompt is processed). A model's need is now its peak. Rows that were not
measured use the same per-token and overhead figures and say `estimated`.
qwen3.6:35b-a3b gets the tested entry for an Apple M5 Pro with 48 GB and
OpenCode, with its measured 67.7 tokens per second, so that row may say
"Tested on a Mac like this".
The ranking changes with it: a pair tested on a Mac like this one comes
first, then speed (ok, unknown, slow), then model rank, harness rank and the
larger context. The fit grade decides whether a row is listed and what it
says, it no longer sorts, so a model is offered at the largest context it
fits at. recommend() takes the budget the runtime measured
(`budgetSource: measured`), and modelFitsBudget() judges one model against it.
Candidate rows now always carry `needSource`, `harnesses`, `installBytes`,
`installGb` and, where one applies, `tag`. `kvBytesPerToken` and the prompt
cache tiers are gone.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- `LocalModelService.launchRefusal(endpointId)`: what a session launch asks
before it starts on the managed endpoint. It answers the error that
refuses the launch (the runtime child is not alive, another program holds
the port, the runtime keeps stopping, the model's digest changed) or
nothing. New error code `runtime-not-running` with its sentence, for a
runtime the person stopped: only a refused session start says it.
- The fake runtime's MacBook log line is rebuilt from the two values the
measurement notes quote, so it is `INFERENCE_COMPUTE_LINE_MACBOOK` now.
Only the Mac mini line is a verbatim capture and keeps `REAL_`.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The types, the module skeletons and the test fake that let the runtime, the
install job and the window be built at the same time.
Types:
- src/types/local-model-install.ts (new): the runtime pin, the install
states, steps and error codes with one sentence and one next step each,
the progress, the `install` object of GET /api/local-model, the bodies and
409 answers of the install verbs, the state file. `LocalModelStatus` moves
here; `install` is optional until the route emits it.
- src/types/local-model.ts: the catalogue row's Mac memory figures, the
measured speed of a tested entry, the `measured` budget source,
`RecommendOptions`, and the row tag, the coding agents and the install
total of a candidate. All optional until the catalogue and `recommend()`
carry them, so the read-only stage keeps compiling and its tests stay
green.
Modules with final signatures and bodies that throw:
runtime.ts (the compiled pin is final), ollama-client.ts, developer-tools.ts,
state-store.ts, install-job.ts (the state, step and transition tables are
final), connect.ts, custom-model-hosts-writer.ts, cli-install-runner.ts.
Nothing imports them yet, so the server is unchanged.
Implemented: paths.ts (every folder of the feature, from dataPath()), and
`modelFitsBudget()` in recommend.ts, the judgement the job makes with the
measured budget before it downloads a model. `recommend()` takes the options
parameter and does not read it yet.
Test fake: test/fixtures/fake-ollama.ts, a runtime on a port the system
picks, in the shapes measured on a Mac: version, tags, ps, the pull stream
(MLX and GGUF lines, an error line, a cut stream, a resumed pull), chat with
and without a tool call, the model list, a runtime archive with Range
support, and a fake child process that logs the device total.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Review fixes for the read-only recommendation and its card.
Recommendation (pure):
- `limited` is now a statement about the Mac: no model with 12 billion
parameters or more fits. It is judged on everything that fits, before
the disk filter and whatever the pick is. A 32 or 36 GB Mac whose pick
is a 9B model no longer reads "This Mac can run a small model" above a
list of 20B, 27B and 35B models, and knowing the chip no longer lowers
the verdict. A small pick under `ok` carries a note that says so.
- A model that fits the memory but has no room on the disk is no longer
dropped in silence: the best-ranked one is named in a note with the GB
that are missing. A full disk cannot turn a 48 GB Mac into a
small-model Mac.
- "Models fit, but not at a context a harness needs" has its own
sentence with the budget and the need, worded as an estimate. The
16 GB Mac mini misses by 1.3 points (10.3 of 10.4 GB) and was told
flatly that its memory is too little, the sentence of an 8 GB Mac.
- The installed memory is named by the size the Mac is sold as (48, not
51.5): the hardware line and the memory sentence. Every computed size
stays decimal GB from the one formatter. An unknown memory size is no
longer reported as "0.0 GB".
- Slow is graded on the unrounded speed: 9.95 to 9.9996 tokens per
second printed as 10 and read ok.
- The disk sentence names the 10 GiB that must stay free, so its
numbers add up for the reader.
Card and Settings group:
- The Settings group "Model on this Mac" is hidden on a server that is
not a Mac, and until the first answer says it is one.
- Only `403 FORBIDDEN` is a refusal. `403 PASSWORD_CHANGE_REQUIRED`
(an admin who must change the password) is a failed request and is
asked again, so that admin no longer loses the card until a reload.
- Each fit and speed label carries what it measures as its tooltip and
accessible name, an alternative with a speed estimate prints it, and
the count beside "Other models that fit" is gone (it was the number
of rows, capped at four, not the number of models that fit).
- The line saying nothing can be set up yet sits right under the pick,
in the colour of every other line (it was 3.9:1), and says whose step
it is.
- The dismissal key is named for this stage, so hiding the read-only
card does not hide the install offer later.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The label list was "frozen for the parallel build" and the card's test said
the route answers 501: both were true while the pieces were written side by
side and are not any more. Comments only.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The read-only face of GET /api/local-model: what the Mac the server runs on
can run as a local coding model.
- local-model-ui.js: one builder paints both containers from the same
payload: the hardware line, the server's headline, the pick with its numbers
(context, download, memory need against the budget, speed), the server's
notes, the other models that fit with their Good, Tight, Slow or Speed
unknown labels, and one line saying that setting it up from Codeman is the
next step and is not built yet.
- The welcome card ships hidden and shows only once the server answered for a
Mac. It stays hidden for not-macos, for a refused request (a non-admin in
multi-user mode) and after "Hide", a per-device localStorage key. The App
Settings group "Model on this Mac" (Models) carries the loading and failure
sentences and is the way back after a Hide.
- Every dynamic string goes in through textContent, both containers are
data-i18n-skip and the module translates its own sentences, repainting on a
language change. Sizes are printed as sent. Labels and harness names come
from the payload: the module names no CLI and the word "Recommended" is not
in it.
- One request a minute at most, shared while in flight; a failed refresh keeps
the last answer on screen.
- i18n.js: zh-CN entries for the module's strings and for every server label.
- frontend-cli-no-id-branching.test.ts scans the new module.
Tests: test/local-model-ui.test.ts (JSDOM over the real markup),
test/local-model-i18n.test.ts.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The read-only route behind the recommendation, and the one module that reads
the CLI registry for it.
- harness-facts.ts: one fact per catalogue harness row, resolved in the CLI
registry at call time: the registry's label, `installed` from the existing
availability probe (asked once per build), `installable` for a stock CLI
whose install command for this platform needs no administrator rights, and
`acceptsModelId`, the test a custom model launch applies to the model id.
A row whose CLI is unknown, disabled or has no custom model mechanism gets
no fact. No CLI id is compared with a literal.
- routes/local-model-routes.ts: answers { hardware, recommendation,
catalogUpdated } in the usual envelope and reformats nothing. Admin only in
multi-user mode (403), checked before anything is probed. A failed probe is
an answer (200, the hardware's `source` says what was read); a 500 is left
for a fault in the catalogue, the harness facts or recommend(). It writes no
file and reads no setting.
Tests: test/local-model-harness-facts.test.ts,
test/routes/local-model-routes.test.ts (app.inject(), no port; the temp HOME
is byte-identical before and after the requests).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The core of the read-only recommendation: what this Mac is, what the catalogue
knows, and which model fits. Nothing is downloaded, installed or written.
- hardware.ts: one bounded async spawn of /usr/sbin/sysctl with seven keys,
without -n, parsed as "key: value" lines by the pure parseSysctl(). Free disk
of the data dir through statfs behind the bounded path probe. Fallbacks from
node:os, a 60 s cache, never throws, never system_profiler. A wired limit of
0 means "not set".
- local-model-catalog.ts: seven models, three harness rows (pi gated) and the
chips table as a compiled constant, validated by a Zod schema on first use.
No override file. Model ids are name:tag in lowercase without a slash; every
model carries paramsB and an (empty) tested list. Chip names match on the
end of the chip string, so 'M4' is not the chip of 'Apple M4 Pro'.
- recommend.ts: device budget (the wired limit when above 0, else 2/3 of
memory below 36 GB and 3/4 from 36 GB, less the reserve), need = weights +
KV cache + 1 GiB + prompt cache, fit good up to 0.85 and tight up to 0.98,
speed ok, unknown or slow (a chip in no chips row is unknown), and the
ranking. Verdicts ok, limited (best model under 12 billion parameters) and
none with its reason. "Recommended" only for a pick with a tested entry for
this Mac, which no row has today.
- types/local-model.ts: comments brought in line with the code (chip matching,
budgetShare rounding, a budget of 0 on an Intel Mac).
- cli-registry-no-id-branching.test.ts: the guard also catches a comparison of
a `cli` field with a CLI id (plus one stray blank line Prettier refused).
Tests: test/local-model-format.test.ts, test/local-model-hardware.test.ts
(the captured output of two real Macs), test/local-model-catalog.test.ts,
test/local-model-recommend.test.ts.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
changesets/action copies a CHANGELOG.md section into the release as it stands,
so a whole release rendered as one bullet behind its commit hash, with
`### Thanks` as literal text and "Minor Changes" as the only real heading.
The Release workflow now rebuilds the body from that section: Thanks first,
then What's new with a heading per change, Fixes, Also in this release and a
compare link, with "Behaviour change:" and "Breaking:" in bold. The release is
titled `Codeman <x.y.z>`; the tag stays `codeman@<x.y.z>`, which is what the
updater and the update toast read.
This cannot be written into CHANGELOG.md instead: `### <Type> Changes` is
hardcoded upstream, and a `## ` line would end the section for changesets/action
and versionSection(). So the changelog keeps its format and the toast is
unaffected. npm has already published when the step runs, so a formatter failure
leaves the changesets text with a warning and the retag still happens.
SECTION_LABELS is now shared with whats-new.mjs so the toast and the page agree
on what counts as a headline. docs/releasing.md records the changeset shape.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Quick-start wrote a selection's launchModel into the session's stored
<Mode>Config, which is what a clear falls back to, so a clear relaunched
the CLI with the endpoint's model id on its native backend. Only API
callers can reach that, since no page sends a clear.
The first launch now takes the forced model from the session's selection,
through the same _withCustomModelLaunchModel() the respawn path uses, so
the stored config keeps the model chosen at create. That also carries the
model onto a new tmux session for a selection applied to a running
session, which the create path used to lose.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The button sat empty in the page while its strip was hidden, which the
mobile accessibility suite counts as a button with no name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
classifyDotGit() checked with lstat and then opened by name, so a FIFO
swapped in between the two blocked a threadpool worker. The open itself
now refuses both, and the handle is judged.
Reported on #772.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Takes the contributor's wording for the picker title and the toast, the
preview refresh after the preflight and the clearer 409 text, and keeps
the fuller tests and the invariants paragraph from the release branch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The poll had no hidden-tab gate, and the local scan's 5 s cache sits under the
8 s interval, so a single tab never hit it: every tick was a full socket sweep
(1.1 s and 772 forks on a box with 212 sockets) for a list nobody was reading.
`_foreignPollTick()` is now the one gate for the first load, the timer and the
return to a visible tab. The `visibilitychange` listener is stored and removed
in `stopForeignPolling()`, since this runs on every trip to the home screen.
Checked in real chromium with document.hidden stubbed: 2 polls in 9.5 s visible,
0 in 17.5 s hidden, 1 on return, and a tab opened in the background makes none
until it is shown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A subfolder whose bounded probe answered `unknown` came back as `children: []`
with no `incomplete` and `truncated` false, so a dead mount, or a healthy one
slower than the 1.5 s probe on first touch, read as an empty folder and a
search missed it in silence. `unknown` is never `absent` (#516).
Such a folder now goes on the walk's `unread` list, which flags the node and
sets `truncated` in both walks. A folder the filesystem says is gone drops out
quietly as before. The row said "file limit reached", which would be false
here, so it now reads "Not fully scanned" for either reason.
Reproduced on a real disk with a parent at mode 0444, where the child's stat
answers EACCES: the API, the search and the panel all report it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
That description is the only warning that the switch weakens Codex's
sandbox, and it had no zh-CN entry. The test runs the real translator over
the real index.html.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
105 new em-dashes sat in comments, test names and one test fixture from the
foreign tmux adoption work. Replaced by hand with commas, colons or
parentheses, one reading at a time. The fixture now quotes the real refusal
text, which says "It may still be running, try again."
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clone Repo's picker now offers New folder, and with the field empty Browse
opens in the cases folder, so the first press there lands on this refusal.
It pointed at Add Case > Create New, which cannot clone. The sentence has to
read right from every picker that offers the button, so it names both tabs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Clone Repo's custom folder became a parent that must already exist (#774),
while its picker still had no New folder button (#761 turned it on for
Create New only). A parent that does not exist yet was refused with no way
to make it from the form, so the same job worked on one tab and dead-ended
on the other.
The test pin that kept Clone off dated from when the field was the exact
target, which could be a folder not yet created. Clone moves to the on side
of that pin, so every picker in the file stays covered.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Edit history panel told an adopted session it needed to be a Claude
session outside Docker, and a stop wait blamed the missing Claude Code hooks
on the mode. Both refusals are right, only the words misled. Each now says
Codeman did not start the session, with a zh-CN entry for the panel.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a foreign session ended, remain-on-exit kept Codeman's own tmux client
on the wrapper's dead pane, so the tab kept its pid and busy, input went
nowhere, both attach routes refused it, and a new session of that name only
led back to the dead tab. paneExit stays out of it by design.
The reconnect watcher's tick now detaches a client from a dead wrapper. That
is an ordinary PTY exit with status 0, so the pid goes null and selecting the
tab runs the attach again: it joins the target if it is back and otherwise
prints the existing guard text. Checked on an isolated instance and in a
headless browser: boot recovery, target gone, target back, view detached, tab
closed, and no restart loop while the tab sits unselected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The remote file reads ran under the user's login shell, quoted for a POSIX
one. Under fish the probe was a parse error, so previews failed, and csh and
tcsh failed every multi-line read. Wrapping the probe in sh -c as the review
proposed would have opened a hole: fish reads \' inside single quotes as a
quote, so nested shell quoting flips, and a path holding ';touch x;' ran
touch x (measured, also on the existing folder listing). csh and tcsh run a
multi-line command line by line, so a newline in a path did the same.
buildRemoteFileCommand now sends every script as one word of printf text
with those bytes in octal, and only sh reads it. Probe, folder listing, read,
range read and transcript lookup gave the same result under bash, zsh, dash,
fish, tcsh and csh in a throwaway container ($SHELL -c, no real sshd), and 24
hostile paths ran nothing. Plain csh caps a word near 8 KB, noted in the docs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The parser and classifier run synchronously on the event loop over text a
pane, a container or an ssh host wrote, and one process could stall the
server for every session: 4 s for a 128 KiB argv of `/pi/` tokens, about a
minute for 2 MiB of socket rows.
The `cli.js` launcher rule leaves the signature regex for a per-word scan,
socket rows dedupe through a Set, the process row regex takes the `s` flag
(a bare CR made `$` fail and `\s+` give its spaces back, quadratic in a row),
argv and pane command are cut at FOREIGN_MAX_ARGV_CHARS, and one matcher per
scan answers each distinct argv once, since every pane of a hostile host can
claim the same process tree.
Measured on a private tmux server with the same 131 KiB argv: the largest
event-loop gap went from 3991 ms to 45 ms. The worst scan the caps allow
takes about 80 ms. Socket rows stay uncapped on purpose (see the docs).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`POST /api/sessions/:id/run` had no adopted gate, so while the wrapper had no
attached PTY it spawned `claude -p --dangerously-skip-permissions <prompt>`
in /tmp, a launch every other adopted route refuses.
The route answers INVALID_INPUT, and `runPrompt()` refuses too, which also
covers the legacy `sendInput()` path and leaves no direct spawn in Session
that an adopted session can reach.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A wrapper that could not start fell back to a direct PTY: `claude
--dangerously-skip-permissions` or a server-account shell in /tmp on this
host, under a tab still carrying the foreign session's name, so someone who
opened an ssh or container session ended up typing into this machine.
startInteractive and startShell now fail the start for an adopted session
and leave `_useMux` on, so selecting the tab tries the wrapper again. The
`error` event carries the reason, since the browser's re-attach does not
read the route's answer. A session with no mux at all is refused the same
way.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A claude session on its own CLAUDE_CONFIG_DIR launched on an endpoint with
<isolated>/projects linked to the server's tree. The resume pin looks there, so
apply found no transcript and started a new conversation, and on clear the
endpoint's transcript stayed behind in the server's tree.
The link now goes to <own dir>/projects when the session set one, so the
conversation and what the endpoint run adds stay on the session's own account
across an apply, a clear and a restart. It goes through linkRealHomeEntry(),
which also creates a missing projects folder instead of leaving a dangling
link, in the own dir and in ~/.claude/projects on the default path. A relative
CLAUDE_CONFIG_DIR of the server gets no link, since joining it would create
projects under the server's working directory.
Rule #516: guardSessionConfigDirHome() now covers configDirVar, so all three
apply sites (route, quick-start, boot recovery) probe the own dir bounded before
the apply touches it, and a dir that does not answer falls back to the server's
tree. Under test, HOME is a temp dir and CLAUDE_CONFIG_DIR is stripped by
test/setup.ts, which a new test pins.
Measured on a real Codeman server with a fake claude: apply and re-apply resume
under the isolated dir with the transcript landing in the own dir, a clear
resumes on the own dir, the isolated dir goes while the transcript stays, and the
same holds across a server restart. The documented limits come out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A remote host whose login shell is tcsh came up on 1.48.0, when the local
sh rewrote ${SHELL:-/bin/sh} to the Codeman host's shell. Since #770 the
launch reaches the far end verbatim and runs tcsh itself, and tcsh refuses
-i -l -c, -l -c and -i -l, so every stock launch left a dead pane.
The login flags are now picked on the far end, in the shape of
loginShellArgs(): set -- -i -l, cleared for csh and tcsh, then
exec "${SHELL:-/bin/sh}" "$@". The text is one brace group, so
cd <path> && { ...; } still starts nothing when the cd fails. bsd-csh takes
every flag combination, csh is listed because it is tcsh on macOS and the BSDs.
The version probe keeps its one statement: the login shell parses it, not
/bin/sh, so fish and tcsh refuse it either way, as they did on 1.48.0.
Measured with the real builder on a real sshd in a throwaway Debian container
(tcsh 6.24.07, bsd-csh, fish 3.6, bash) and through a fake sshd with
tcsh 6.24.10, fish, zsh, dash and bash, a real Codeman server on tcsh included.
The tcsh disclosure comes out of the docs and the changeset. They keep the two
limits 1.48.0 had too: tcsh has no export, so the PATH bootstrap only prints an
error, and the tmux probe needs a command executable, which Debian lacks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The attach line now takes the rule #770 gave remote and docker launches: bashCArg() single-quotes it at both respawn-pane sites, so it reaches bash verbatim. quoteAdoptLaunch is retired, since its double quotes kept the same $ and backtick safety but lost a real newline. #510's no-cd condition stays.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The "Will clone into" line named a folder the request would not use: the
preflight replaces the page's guessed case name with the server's (for a
URL such as .../repo?tab=readme the guess is repo-tab-readme and the server
says repo) and nothing refreshed the preview, so it stayed on the old name
while the field and the request moved on. It is now refreshed there too.
The Browse picker's title and the empty-field toast still said "Choose the
folder to clone into", the old exact-target meaning. Both now ask for the
parent folder, matching the field's label.
CLAUDE.md and the invariants paragraph split the two meanings: path on the
API is the exact folder, the Clone Repo tab's field is a parent and the page
sends <folder>/<case name>.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`_renderMarkdown` wrote dir="auto" into its string, before the code pins and
the linkifier. An element carrying dir="auto" re-resolves its direction on
every insertion under it, and a block whose letters all sit in `dir` children
walks its whole subtree each time, so a code-only or path-only document went
quadratic, English ones included.
Measured in real Chromium through the real File Viewer, before and after:
9000 inline-code items 8.7 s to 0.35 s, 9000 bare paths 5.6 s to 0.23 s, a
209 KB out-of-workspace list 41.9 s to 1.4 s. All three now match master.
The pass is now `_applyBlockDirection()`, run last on the DOM by both
consumers. The direction results are unchanged. Tests pin that the rendered
string holds no dir="auto" and that the pass runs after the linkifier.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
startShell() writes `clear\n` 100 ms after a new mux attach, and for an adopted
session that write reaches someone else's live pane through the wrapper's tmux
client. Measured on an isolated instance: a shell with `echo DRAFT_NOT_SUBMITTED_`
typed at its prompt ran `echo DRAFT_NOT_SUBMITTED_clear` the moment the tab
opened. `shell` is the feature's primary case and also the classifier's answer
for anything it cannot name (vim, psql, a `read` prompt), so every adoption put
those six bytes in front of whatever was reading stdin.
After the fix a byte recorder in the foreign pane saw 0 bytes in the 6 s after
adopting, and a real input afterwards arrived intact. Not in the review: its
recorder covered the trust-dialog path, which is `claude` mode.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codeman maintainer