Doctor and repair
benchbar doctor # read only, exit 1 when something failsbenchbar repair --dry-run # show the plan, change nothingbenchbar repair # apply, with a confirmation promptDoctor is read only, with one exception: the one time move of the state
folder on the first run after an upgrade (~/.local/share/benchbar/.benchbar
to ~/.local/state/benchbar), which renames and links, and never copies
or deletes.
A few names used below: the launchd agent is the macOS background
job that runs the bench; honcho is the process manager that starts
the processes listed in Procfile.lean; the runner is the
script the agent starts; the stop flag is the file that says
whether the bench was stopped on purpose (manual) or paused after
crashes (crash), so it is not restarted.
Doctor checks the Homebrew formulae, the bench env, bench version, the
socket.io module, the built assets that assets.json references, honcho,
Procfile.lean, the runner, the launchd agent and its last exit code,
the stop flag, the shell helpers, legacy agents, the MariaDB bind address
and charset config, the PDF engine (wkhtmltopdf, and on v16 the Chromium
frappe can use), a stray Homebrew Redis on 6379, Full Disk Access for
crontab, the toolchain as the bench sees it (Node, yarn, the MariaDB
server, pkg-config), honcho without pkg_resources, the fork safety
variables, stale processes on the bench’s ports, the site
ping, whether the running bench serves every site, /etc/hosts, log sizes, CleanMyMac, Mole, and port clashes with
other benches. Every warning and failure names its fix.
Repair runs only the flagged fixes, in dependency order, and shows the
plan before it asks. A broken env/ is moved to
env.broken.<timestamp>, never deleted. When every check passes, repair
prints unchanged: all N checks pass, nothing to do, and a second run
after a repair should say the same. The most common case,
a cleanup tool that removed env/, node_modules and the built assets,
is covered in Troubleshooting.
Reading the output
Each check prints one line, [OK], [WARN] or [FAIL], with its label
and a message. A fix: line with the exact command follows every WARN
and FAIL. benchbar doctor --json prints the same checks with their id,
level, fix_command and action: the id is the heading of each check
below, and the action is what repair would run. The format is in the
JSON schema.
Doctor exits 1 when any check fails. Warnings alone exit 0. It never changes anything in a bench, never asks for a password and never reads the Keychain, so the app runs it on a timer.
Three flags change what the report holds:
--prerequisiteschecks a Mac before an install, without a bench: the first run wizard in the app shows the same list.--quickskips the checks that start brew, python, node, bench or git, or ask the site, and names them.--jsonlists them inskipped_checks.--ack CHECKsilences one warning you have seen. It writes one line to the bench’s state file with the message it silenced, so the warning returns when its message changes.--unack CHECKundoes it. A failure cannot be silenced, andrepairignores silenced warnings.
Checks with a repair action are fixed by benchbar repair. The others
name a command for you to run, because the change is a person’s call
(for example which branch an app follows) or needs another app (Full
Disk Access, CleanMyMac, Mole). Commands and flags are in the
CLI reference.
System checks
The checks run in four groups, in this order: system, bench, service and
site. benchbar service and adopt look only at the service group.
brew
The Homebrew formulae of the bench’s profile are installed: the Python,
Node and MariaDB formulae, and redis. A missing one fails. The build
formulae pkgconf and mariadb-connector-c (needed to build
mysqlclient for frappe v16) are a warning when missing.
Fix: install Homebrew from https://brew.sh when it is missing; run
00-mac-system-deps.sh --profile <profile> from the checkout for missing
formulae, or brew install pkgconf mariadb-connector-c for the build
formulae. When only the profile’s Node formula is missing (the profile
moved to a newer Node), benchbar repair installs it.
formula_dates
Homebrew deprecates a formula about a year before it disables it, and a
disabled formula can no longer be installed, so a fresh install and every
brew install fix of the profile fail from that day. This check reads the
dates of the profile’s Python, Node and MariaDB formulae from the local
tap (brew info --json, no network): a WARN within 90 days of a disable
date, a FAIL once it has passed or once Homebrew no longer knows the
formula.
Fix: update BenchBar (brew upgrade benchbar, or Check for Updates in the
app); a newer profile names the current formulae, and benchbar repair
then moves the bench to them. A bench keeps the formulae it has.
python_leaves
The profile’s Python (for example python@3.11) exists, has the version
the profile expects, and was installed on request, so brew autoremove
cannot delete it.
Fix: brew install or brew reinstall the formula when it is missing or
the wrong version. When it is only a dependency, repair marks it installed
on request (brew tab --installed-on-request <formula>).
mariadb_bind
MariaDB listens on 127.0.0.1 only, or, when it is not running, the
bind address drop-in frappe-mac-local-only.cnf exists. A server
reachable from the network is a warning.
Fix: benchbar repair writes the drop-in into
$(brew --prefix)/etc/my.cnf.d/ and restarts MariaDB when it is
running. Every bench on this Mac shares that server, so running benches
lose their database for a few seconds; stop them first, or repair when
nothing is mid request.
mariadb_utf8
The utf8mb4 drop-in frappe.cnf in $(brew --prefix)/etc/my.cnf.d/ is
current, and my.cnf includes that folder. Frappe needs utf8mb4 server
wide. A drop-in written by someone else that sets utf8mb4 is accepted and
left alone. Doctor reads the files only; it never logs in to MariaDB.
Fix: benchbar repair writes the drop-in and the !includedir line,
then restarts MariaDB when it is running and the drop-in changed. As
with mariadb_bind, every running bench loses its database for a few
seconds.
pdf_engine
wkhtmltopdf is the patched Qt build, the one Frappe needs for PDFs, and
no unpatched Homebrew wkhtmltopdf comes first on PATH. On a v16
profile it also says whether the Chromium that Print Formats set to
“chrome” use has been downloaded.
Fix: benchbar repair installs the official patched package (with
sudo). A shadowing Homebrew build: brew uninstall wkhtmltopdf. The v16
Chromium: cd <bench> && bench setup-chrome.
redis_6379
Nothing listens on port 6379. The bench runs its own Redis on its port block, so a Homebrew Redis on 6379 is unused.
Fix: brew services stop redis, only if nothing else on this Mac needs
it. benchbar repair offers the same and asks first; under --yes it
leaves Redis running and prints the command.
wsl_time_sync
Linux in WSL only; it is left out elsewhere. Two clocks set the time of a
WSL distro: the Windows host (Hyper-V time sync) and NTP inside the distro
(systemd-timesyncd). Windows syncs its own clock about once a week; while
it is off from NTP by more than a fraction of a second, every NTP poll
(about every 32 seconds) steps the clock back. A step can end the worker’s
blocking Redis wait, and then the whole bench stops and restarts. The host
sync cannot be turned off from inside WSL, so NTP in the distro is the one
to turn off. The check is OK when timedatectl show -p NTP --value says
no or timedatectl is missing, and a warning when it says yes.
Fix: sudo timedatectl set-ntp false. Repair does not run it, because
sudo on Linux is kept for apt, the wkhtmltopdf package and MariaDB.
cleanmymac
CleanMyMac is not installed in /Applications or ~/Applications, or in
the Setapp folder inside either. Its cleanup can delete env/,
node_modules and public/dist.
Fix: in CleanMyMac, add the bench folder to the Ignore List before running any cleanup.
mole
Mole (mo) is not installed, or the bench is protected in its whitelist.
mo purge looks for projects under ~/dev and similar folders and
deletes their node_modules and dist folders and any folder with a
CACHEDIR.TAG, which includes the bench’s env/. It skips any path in
~/.config/mole/whitelist and everything below it, so the check passes
when a line there is the bench folder or a folder above it. Lines with
*, ? or [ are not read.
Fix: when ~/.config/mole/whitelist does not exist yet, first run
mo clean --whitelist and save once: a whitelist file replaces Mole’s
built in entries, and its editor writes them into the new file. Then
add the bench folder, with your bench’s path in place of <bench>:
echo '<bench>' >> ~/.config/mole/whitelist.
app_copies
At most one BenchBar.app, in /Applications or ~/Applications. Two
copies, typically Homebrew’s cask in /Applications and the one line
installer’s in ~/Applications, share one bundle id, so macOS may open
either, and the login item and the updater may each pick a different one.
Fix: quit BenchBar and move the copy you do not use to the Trash. With
Homebrew’s cask benchbar-app installed, that is the one in
~/Applications. benchbar repair never deletes an app.
full_disk_access
crontab is readable. Without Full Disk Access for the Terminal, bench init and bench setup backups fail with “Operation not permitted”. Not
checked when BenchBar runs doctor, since the permission belongs to the
Terminal that runs bench.
Fix: System Settings > Privacy & Security > Full Disk Access: add Terminal (or the app that runs benchbar), then open a new window.
Bench checks
env_python
env/bin/python exists, runs, and is the Python version the profile
expects.
Fix: stop the bench first (benchdown), then benchbar repair rebuilds
the env; the old one is moved to env.broken.<timestamp>. Repair
refuses to rebuild the env while the bench runs. When no profile matches the bench’s Frappe (a
develop or v17 bench with no stored profile: doctor’s header then says
profile v15-lts (default: no profile matches this bench)), the Python
to rebuild with is unknown, so a version mismatch is only a warning, no
rebuild is offered, and the fix is benchbar install --profile NAME.
env_setuptools
On Frappe v15, env/bin/python imports pkg_resources: bench and honcho
need it, setuptools 70 and later dropped it, and an env made with Python
3.12 or newer has no setuptools at all. Skipped on other versions.
Fix: benchbar repair installs setuptools<70 into the env (an env
rebuild on v15 does the same).
bench_version
bench version works in the bench folder. The message also says whether
pipx or uv owns bench. A failure is classified first: a missing bench
command, or one whose own venv is broken (exit 126 or 127), is the CLI’s
problem and gets uv tool install frappe-bench or the pipx equivalent;
an app in sites/apps.txt that does not import is named, with bench setup requirements --python as the fix. Neither moves the env aside.
Fix: benchbar repair rebuilds the env only when frappe itself does not
import and the profile is this bench’s (stored or detected).
toolchain_node
The node the bench’s processes see (launchd’s PATH, not your shell’s)
has the major version the profile expects. nvm’s node is only on your
shell’s PATH, so the bench does not see it.
Fix: brew install <node formula>; the bench’s PATH puts it first.
When the profile’s formula is not installed (a profile that moved to a
newer Node), benchbar repair runs the install (node_install); the old
formula is never removed.
toolchain_yarn
yarn is on the bench’s PATH. bench build needs it.
Fix: npm install -g yarn with the bench’s npm; benchbar repair runs
it (yarn_install), since yarn is global to one node formula.
mariadb_version
A MariaDB server listens on 3306 and its version is in the range the profile accepts: 10.6 to 10.11 for v15, 10.6 to 11.8 for v16. Outside the range is a warning, as it is in Frappe.
Fix: brew services start <mariadb formula> when nothing listens;
brew install <mariadb formula> for a version outside the range.
toolchain_pkgconfig
pkg-config is on the bench’s PATH and finds mariadb-connector-c.
mysqlclient for frappe v16 needs both.
Fix: brew install pkgconf mariadb-connector-c.
socketio
apps/frappe/node_modules/socket.io exists. Without it socketio fails
with “Cannot find module ‘socket.io’”.
Fix: benchbar repair runs bench setup requirements --node.
assets
sites/assets/assets.json exists and every dist file it references is
there. Missing files make the site load without styling.
Fix: benchbar repair runs bench build.
apps_txt
Every app in sites/apps.txt has a folder in apps/, and every git app
in apps/ is listed. A listed app without a folder fails, because every
bench command then fails to import it. An unlisted git app is a warning:
usually a get-app that did not finish.
Fix: benchbar app add <app> for a missing app. For an unlisted one, move
the folder aside, then benchbar app add <its git URL>. Repair has no
action here: changing code is your call.
app_branch_policy
Each git app is on the branch config/apps.tsv names for the profile.
Fix: git fetch and git checkout the policy branch in the app, only if
you meant to follow the policy. Repair has no action here.
app_update_pending
One warning for each benchbar app update that stopped after it had
started changing things (a failed backup, requirements, migrate, build or
restart). The row names the step it stopped at, why, and how many steps
are left. Nothing is wrong with the bench’s files: the update is simply
not finished, and an app that is already at its new commit would otherwise
look up to date while its migrate never ran.
Fix: run benchbar app update <app> again. It runs the steps that did
not finish and nothing else, then removes the record. --discard-pending
drops the record and stops: no step and no update runs, so run the update
again afterwards if you still want it. An unreadable (cut short) record is
kept, nothing runs from it, and doctor says so. Doctor reads the record only, no
git and no bench. Repair has no action here.
app_add_pending
The same for the planned benchbar app add (--dry-run --json, then
--apply): a build, an install or a restart that did not finish. A failed
build also skips the install, so the app is not installed on a site whose
assets were never built.
Fix: benchbar app add <app> again. The next plan lists the steps that
are left, with the reason, and applying it runs them. Repair has no action
here.
dependency_behind
One warning for each app a focus app needs (its required_apps in
hooks.py, followed through other apps) that is behind its remote
branch, for example exponent_custom_v1 (needed by exponent_ecr) is 30 commits / 12 days behind upstream/develop. Commits and days follow the
remote branch itself (git’s first parent line): the commits it got after
yours, and the age of the first of them. What a merge brought in from
other branches or older history does not count, so a release merged in
is one commit. A focus app itself never gets this warning: you pull the
app you work on yourself. See
focus apps.
Doctor stays read only and never fetches on its own: the numbers come
from the remote branches as git last fetched them, your own git fetch
included. When benchbar has not fetched them in the last day, the row
says so (“as of a fetch 3 days ago”) and the OK line says to run
benchbar doctor --fetch. That fetches the focus apps’ dependencies
first, with a 20 second timeout each and never a password prompt; only
.git changes. It never runs with OFFLINE=1 or --dry-run. An app
never fetched is reported as unknown, and a fetch that fails is no
failure either.
Fix: benchbar app update <app>, which shows the changelog, backs up
the sites that have the app, fast forwards it and migrates. A dependency
with local changes gets git status first. Never bench update.
apps_behind
One line about every other app: neither a focus app nor needed by one.
It reads the remote branches as they were last fetched (even doctor --fetch does not fetch these) and is always OK.
lock_parse
The bench’s benchbar.toml lockfile, when one is set, exists and parses.
See the team lockfile.
Fix: correct the file by hand, or write it again with benchbar lock write. A lockfile path that does not exist yet: benchbar lock write --lock <path>.
lock_drift
The bench matches its lockfile: the same apps, repos, branches and pinned commits. Drift is a warning.
Fix: benchbar lock apply (benchbar lock check lists the differences).
profile_outdated
The bench’s team profile comes from a subscription whose clone is behind
its remote, as of the clone’s last fetch. Doctor itself never fetches;
benchbar doctor --fetch fetches the subscription first (not with
BENCHBAR_OFFLINE=1), and the message says when the last fetch is
older than a day. See
Sharing a profile.
Fix: benchbar profile update NAME (shows the changes and asks).
logs
bench.log, worker.log and worker.error.log are under 50 MB each.
Fix: benchbar repair copies each large log to <name>.old.<stamp> and
empties the live file in place, so the bench’s processes keep writing to
the file they have open. At most three copies are kept.
Service checks
port_block
Shown only in the plan of install, adopt or service when the bench’s
ports are about to move to another block (because another bench uses
them, or --port-offset asked for it). Doctor never shows it.
Fix: part of the plan: benchbar service --port-offset <n> moves the
ports with bench set-config -g.
bench_path
The bench path and the default site name are plain text to the runner
script (BENCH="..."), the launchd agent’s plist and a hosts line: no
double quote, backslash, $, backtick, <, >, & or control
character, and no whitespace in the site name. Spaces and apostrophes in
the path are fine. Doctor only reports it; service, repair, adopt,
install, up, autostart and site add refuse to write for such a
bench and name the character.
Fix: move the bench to a folder without that character, then
benchbar adopt <new path>; for a site name, benchbar site default <name> or rename the site folder.
honcho
honcho, which runs the Procfile, is found on PATH, in the pipx venv of
frappe-bench, or in the bench’s env/bin.
Fix: benchbar repair installs it into the bench env. adopt never
installs into env; install it with pipx install honcho first.
honcho_setuptools
honcho imports cleanly. honcho 1.x imports pkg_resources, which Python
3.12 and later only have with setuptools installed.
Fix: benchbar repair installs setuptools into honcho’s own venv only.
procfile
Procfile.lean exists, was written by benchbar, and matches its
template and the bench’s settings (for example the scheduler choice).
Fix: benchbar repair writes it again; the old copy goes to the
backups.
runner
The runner script benchbar-run.sh is current, and no old
frappe-mac-run.sh is left in the bench.
Fix: benchbar repair writes it again. A running bench picks the new
runner up on its next start.
agent
The launchd agent plist is current and loaded, and its last exit code is not an error.
Fix: benchbar repair writes and loads it. A non zero last exit code:
benchbar logs shows why.
runner_heartbeat
A running runner rewrites logs/.benchbar/heartbeat every 30 seconds
(runners written by BenchBar 0.6.1 and later), and the app trusts the
runner’s state.json only while that file is under 90 seconds old. A
warning means the bench runs a runner process older than its script:
benchbar service or repair rewrote the script, but a running bench
keeps the process it started with.
Fix: benchbar restart. Until then the app checks that bench with the
CLI once a minute, as it did before 0.6.1. When the script itself is
outdated, the runner check says so and this one waits.
fork_safety
The agent passes OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES and
NO_PROXY=* to every process. Without them macOS kills forked workers.
Fix: benchbar repair writes the agent again.
scheduler
Reports whether the scheduler runs in Procfile.lean. It is a choice, so
this check never warns.
Fix: none needed. benchbar service --with-schedule turns it on,
--without-schedule off.
share_scheduler
Warns when a benchbar share setup stopped after it paused the scheduler of
every site on the bench (pause_scheduler in common_site_config.json) and
did not resume it. The restore of the share copy needs that bench wide
pause, so setup leaves a marker in its state folder first. A pause without
the marker is yours and is never flagged.
Fix: cd BENCH && bench config remove-common-config pause_scheduler, or
benchbar repair. The next share setup or share start resumes it too.
stop_flag
No stop flag, or a stop on purpose (benchdown). A flag of crash (three
crashes in ten minutes), broken (honcho or env was missing) or
port_conflict (a process that is not this bench’s held one of its
ports when the runner started) means auto restart is paused.
Fix: benchbar logs, fix the cause, then benchup. For broken:
benchbar repair, then benchup. For port_conflict: benchbar doctor names the process; free the port or run benchbar ports setup,
then benchup.
helpers
The # >>> benchbar >>> block in ~/.zshrc ($ZDOTDIR/.zshrc when
ZDOTDIR is set) with benchup and the other helpers is present and
current. Old helper blocks from earlier setups are
reported. A block whose BENCHBAR path is gone fails: every helper runs
that path. That happens after brew cleanup deleted a versioned Cellar
folder an old block named, or after the one line installer’s checkout
went to the Trash.
Under Homebrew a block that runs a git checkout with its own
.benchbar folder is a warning and stays as it is: Homebrew’s CLI never
reads that checkout’s state, so benchup would lose its benches. A
block that names Homebrew’s or the one line installer’s benchbar
stays as it is when that copy hands off to the CLI inside BenchBar.app.
The app’s CLI writes a block that names Homebrew’s or the installer’s
benchbar when one of them hands off to it (the copy that was run first,
then Homebrew’s, then the installer’s), and the app’s own link in
~/.local/state/benchbar/bin only when nothing else is installed: that
link goes with the app, and benchup would go with it, while a Homebrew
or installer copy runs the app’s CLI as long as the app is there and
itself once it is gone. A block that names the link while such a copy is
installed is outdated (0.7.3).
Fix: benchbar repair writes the block. Remove an old block by hand.
cli_link
~/.local/bin/benchbar and ~/.local/bin/frappe-mac are links to this
checkout, so both names work from any folder. When Homebrew has the
benchbar formula, its own bin folder puts benchbar on PATH: the
links are optional then, and a link to Homebrew’s benchbar always counts
as current. Under Homebrew a link that leads anywhere else, such as the
one line installer’s checkout or a Cellar folder, comes first on PATH and
is a warning. A link to a git checkout with its own .benchbar folder is
a warning too, but repair leaves it: that checkout’s state is not
Homebrew’s. For the CLI inside BenchBar.app, links to a Homebrew or
installer copy that hands off to it are current.
Fix: benchbar repair writes the links. With Homebrew’s benchbar
installed it never makes a new one: it points a wrong link at this CLI
and keeps the old link in its backups. A file that is not a link is left
alone: move it aside (and link again, without Homebrew).
cli_duplicate
One copy of the CLI: not the one line installer’s in
~/.local/share/benchbar next to Homebrew’s benchbar formula. Both
copies work on the same state folder and the same benches, but only one
of them is benchbar on PATH, and the other one’s repair or
self-update changes the copy nobody runs. A git checkout of your own
warns too while Homebrew has the formula. It keeps its own state, which
Homebrew’s CLI never reads, so for a checkout the fix is brew uninstall benchbar.
Fix: to keep Homebrew’s, run its repair by its full path,
$(brew --prefix)/opt/benchbar/bin/benchbar repair, then move
~/.local/share/benchbar to the Trash; the benches and the state stay.
The check offers the Trash only once the state has moved out of that
folder: while a run of the old CLI holds its lock, run repair again when
it ends. When the two folders are on different volumes the state stays
in ~/.local/share/benchbar/.benchbar, so keep that folder. To keep the
installer’s, brew uninstall benchbar. There is no repair action: which
copy stays is your call.
The CLI inside BenchBar.app (0.7.1) next to Homebrew’s or the
installer’s copy is the normal case: since 0.7.1 those hand off to it,
so this check passes. A copy too old to hand off (before 0.7.1) is a
warning, because benchbar in Terminal may run that older copy. Fix:
brew upgrade askysh/tap/benchbar for Homebrew’s, git -C ~/.local/share/benchbar pull --ff-only for the installer’s. A copy newer
than the app’s CLI (0.7.3) is a warning too: it hands off to the older
one, so every benchbar runs the older version until the app updates.
Fix: Check for Updates in BenchBar, or brew upgrade askysh/tap/benchbar-app for the cask’s app.
From the app’s CLI the check also looks for a second state folder (0.7.3):
a real .benchbar (or .frappe-local) in ~/.local/share/benchbar next
to the state in ~/.local/state/benchbar means two states that never
merge, left by an older CLI or by a layout the move never saw. Fix: move
anything you still need out of it, then move the folder (the installer’s
whole checkout when its benchbar is still there) to the Trash. When the
new path is a link to that folder (the two are on different volumes) it
holds the state and the check says to keep it. When the move was cut
short (the old folder holds the app’s bin/, or the new folder holds no
state yet) the check says so and the next benchbar run completes it;
nothing is offered for the Trash then. Nothing is ever deleted for you.
legacy_agents
No per process LaunchAgents (one each for web, worker, socketio and so on) and no agent from frappe-mac 0.2.
Fix: benchbar repair boots them out and moves the plists to
~/Library/LaunchAgents-disabled/<timestamp>/.
dead_agents
Every loaded com.benchbar.* agent still has its runner script. When a
bench folder is emptied or deleted (a cleanup tool, rm) while its agent
stays loaded, launchd starts the missing script every 20 seconds, it
exits with code 127, and each try adds a line to the bench’s log. The
check looks at every bench’s agent, not only the current one.
Fix: benchbar uninstall-service --bench-dir <path>, with the path the
warning shows. It works when the folder is no longer a bench: it boots
out the agent that points at the path and moves its plist to
~/Library/LaunchAgents-disabled/<timestamp>/.
hosts
/etc/hosts maps the site to 127.0.0.1.
Fix: benchbar repair adds the line inside its /etc/hosts block (with
sudo). A declined question leaves the step skipped.
port_clash
No other bench benchbar knows uses this bench’s web, socketio or Redis ports, running or not. Two benches set up with the same ports cannot run at the same time.
Fix: benchbar service --port-offset <n> --bench-dir <bench> moves this
bench to the next free block (or stop the other bench). See
Port blocks.
orphans
No stale process holds the bench’s ports while the agent and honcho are
not running. A killed bench start can leave Redis, socketio or gunicorn
behind, and the agent then cannot start.
Fix: benchbar down (or benchbar restart) stops only this bench’s
leftovers.
Site checks
ping
http://<site>:<port>/api/method/ping returns 200. A stopped bench is
fine; processes that run but do not answer fail.
Fix: benchbar logs shows why; benchup starts a stopped bench.
site_not_served
Every site folder in sites/ is served by the running web process. The
check asks each site once with its own Host header; a 404 means the
folder is on the bench but the process does not serve it (it was started
before the folder appeared, or something else holds the port). It asks
only while the bench runs: a stopped bench is never pinged or woken, and
the check passes with “the bench is not running”.
Fix: benchbar restart. benchbar status --json shows sites[].served
(true, false, or null when the bench is stopped or did not answer).