New [gaming_budget] block reads steam-backlog-enforcer's /api/budget every
30s: "🎮 2h00 left · 6h00/8h00 +📖1h". Colours follow the enforcer warn_at
(1h/30m/10m), blocked is red, a dead server shows the cached value as stale.
Unearned bonus sizes come from the enforcer config via jq --slurpfile.
Prunes the oldest evidence artifact to stay under the 60-file cap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CD7nSVE2dKvByG8hdSi3gm
Android runs one instrumentation per package: a second `am instrument`
force-stops the first ("stop com.kuhy.a11ydump due to start instr"), and
so does `install -r`. The live two-session run on the Pixel caught two
sessions' dumps killing each other ("Process crashed"). dump_display now
runs install + instrument under a per-serial flock; a dump takes well
under a second, so queueing costs nothing noticeable.
After the fix the same 300 s run had 0 failures across 659 actions, and
the real screen and both apps stayed where they were.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpA6wW2DKiovQADTdjWJoH
A raw `adb shell input tap` took no lease, so one session's tap landed in
another session's app (2026-09-25). python_pkg.phone_guard classifies every
adb call in a Bash command and, as the Claude Code PreToolUse hook, takes
the matching phone_lease (one app / real screen / whole phone) or exits 2:
key/text input without -d, input on another session's virtual display, or
a lease still held elsewhere after 40 s.
android_ui gains --display self|<id>: it drives only this session's
phone_vd display under that app's lease, and every input carries -d.
uiautomator sees display 0 only, so virtual-display trees come from a tiny
instrumentation APK (A11yDump.java) built from source without Gradle.
Prunes the oldest evidence artifact to stay under the 60-file cap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WpA6wW2DKiovQADTdjWJoH
Several sessions can now share one phone: a lease covers one package (the
real screen stays the default scope, so existing callers keep their
meaning), and only phone-wide actions take `--whole-phone`, which is capped
at 300 s and runs its `--cleanup` commands exactly once on release or, via a
detached reaper, on expiry. The serial is always resolved, never the
"default" placeholder that let android_ui and phone_deploy lease different
keys for the same phone (2026-09-25). Adds status --json, holds and whoami
for phone_vd.sh and the raw-adb hook.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012Ajzz98xksbfak8CaunZWC
Filled scalloped canopy over a stroked J-handle (filled because a stroked
dome with ribs reads as a jellyfish at 48dp). AppIcon moves to _app_icon.py,
re-exported from apps.py, so the registry stays under the 250-line cap.
Prunes the two oldest evidence artifacts (60-file cap; still in history).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012Ajzz98xksbfak8CaunZWC
Two sessions driving the same phone interleave their taps: on 2026-09-20 a
tap meant for the workout sandbox opened another session's manga reader,
twice. The lease is a flock-guarded JSON file per serial under
~/.cache/phone-lease/, owned by the nearest ancestor `claude` process (one per
session), refreshed on every android_ui call and expiring 180 s after the
last one, so a quiet session frees the phone without anyone remembering to.
A foreign fresh lease makes a call wait up to 60 s, then exit 3 naming the
holder and pointing at the alternatives (the phone's virtual display, the
headless emulator) instead of inviting a retry loop. phone_deploy.sh takes
the same lease (exit 40).
The oldest evidence artifact is pruned to stay under the 60-file cap; it
remains in history.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YbBZqUaNWyhHYfzz6zGC7q
`secret-tool lookup` is non-interactive — on a locked keyring it returns
nothing and shows no dialog — and a busctl Unlock+Prompt pair cannot work
either, because Secret Service prompt objects die with the D-Bus
connection that created them. secretstorage keeps one connection:
get_default_collection().unlock() pops the gcr dialog and blocks until it
is answered. Verified on the real profile: 5 v11 rows -> v10, then with
the keyring re-locked Helium (through the wrapper, so the flags file and
LeechBlock apply) lists all 5 logins and keeps every row.
`pgrep -f /opt/helium-browser-bin` also matched the invoking shell when
its argv mentioned the path, so the guard refused on an idle machine;
now anchored to `^/opt/helium-browser-bin/`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8YGXVZdNUG79q5DGsdMm3
Chromium 151 removed the last Manifest V2 code paths and the Web Store
purged uBlock Origin on 2026-08-31, so ungoogled-chromium 153 cannot load
the full blocker any more. Helium (GPL-3, ungoogled-chromium base) bundles
uBO 1.75.0 as a component extension and keeps MV2 alive.
- installers/install_helium.sh: pacman deps, AUR via yay or makepkg
(+validpgpkeys import), /usr/local/bin/helium-browser ->
browser-preexec-wrapper so LeechBlock NG and the hosts refresh still
apply, --clone-profile with a downgrade check and running-browser
refusal, then the verifier. Passed in vmbox from scratch, then host.
- installers/verify_helium.py: Xvfb + DevTools exit-code gate — uBO
background_page + manifest_version 2, every /etc/chromium/policies key
on chrome://policy (Helium reads that dir natively), adblock-tester >= 95.
- installers/migrate_helium_passwords_to_basic.py: lightdm autologin
leaves gnome-keyring locked, so the 5 v11 logins make Helium prompt on
every launch; this re-encrypts them as v10 and only then enables
--password-store=basic.
- register helium-browser in aur_packages.txt and the periodic browser
wrapper/service lists; pylint hook gets websockets + cryptography;
codespell learns the vendor name; anthropic 1.6.0 -> 1.7.0 (gate red);
oldest evidence artifact pruned to stay under the 60-file cap.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01M8YGXVZdNUG79q5DGsdMm3
kuhy: "I do not want it to work like this, I want it to just work
normally." PS4 mode (046D:C260) cannot be supported on this PC (no kernel
driver, SDL lg4ff has no c260 either, so FFB is dead there too) and the
selector is physical. Keep the mode detection so the journal and --status
name PS4 mode instead of "not found"; drop the desktop notification and
the 5-minute Restart loop.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NVfC2T6DyvCFqse9A99zef
On the 2026-09-15 and 2026-09-17 boots the wheel enumerated as 046D:C260
(PS3/PS4 selector in PS4 position, bound by hid-generic). The BPF udev
rule, the ensure unit and its hotplug rule all keyed on C24F, so the unit
failed once with "no G29 found" and nothing was said; 1st/3rd stayed on
the firmware's bad decode until the selector was flipped a day later.
- fixes/g29_wheel_mode.sh (new, sourced): classify the HID bus as
native/ps4/compat/absent; root-safe desktop notification via
/run/user/*/bus + sudo -u.
- --ensure: absent -> exit 0; ps4/compat -> exit 1 + critical
notification; native -> attach. --status prints the mode first.
- g29-shifter-bpf.service: Restart=on-failure, RestartSec=5min so a
notification lost before the session bus exists at boot is re-sent.
- 99-g29-shifter-ensure.rules: also fires on C260.
- tests/test_fix_g29_shifter.bats: 13 tests against a fake sysfs tree.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NVfC2T6DyvCFqse9A99zef
The dkms pacman hook already rebuilds every preset on install and remove,
so the unconditional mkinitcpio -P cost 20 s twice per run on the host.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SS8M5wk88nJTXxDWz61KKk
Live on this host since 2026-09-11 but never committed. The wheel firmware
decodes the shifter's hall sensor itself and this unit's top row rests at
Y 180-193, below the ~195 engage threshold, so 1st/3rd often did not
register. g29_shifter.bpf.c re-decodes those two gates from the raw X/Y
bytes with hysteresis (engage >= 160, release <= 140) and ORs the bit in
before hid-logitech parses the report; it never clears a bit, so the
firmware's working gates cannot regress. fix_g29_shifter.sh builds,
attaches and installs it persistently (udev-hid-bpf rule + a systemd
safety-net unit pulled in on every re-enumeration); g29_shifter_capture.py
records and grades raw captures per gate.
Two files were over the 250-line cap: the test file is split at the
record/CLI boundary with the report builders moved into a conftest fixture
(the naming hook allows only test_*.py in tests/), and the grader's
docstrings are tightened. No behaviour change to the installed fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SS8M5wk88nJTXxDWz61KKk
The "flags nothing on a quiet machine" case read the real /proc/loadavg and
went red whenever a game was running on the host (22/16 during a BeamNG
session), and the "flags a load at/above thread count" case had to SKIP
whenever the host was idle. Same override pattern as JOURNALD_CONF_DIR: the
harness writes a fixture, both cases now assert deterministically, and a
missing file degrades to 0 as before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SS8M5wk88nJTXxDWz61KKk
BeamNG.drive 0.39 native lost G29 FFB because SDL3's HIDAPI lg4ff driver
advertises an effect-status query it cannot answer (12,201 failures per
session in beamng.log). DOCS-g29-ffb.md records the diagnosis, the narrow
SDL_JOYSTICK_HIDAPI_LG4FF=0 hint vs the broad one, and the launch-option
%command% bug that had gamemoderun passed as a game argument.
install_new_lg4ff.sh swaps the in-tree hid-logitech for new-lg4ff via DKMS
(dkms.conf installs it under the same name in updates/dkms/), regenerates
initramfs presets that bundle the module, and offers --status/--revert.
Passed install/status/revert in vmbox; host run pending.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SS8M5wk88nJTXxDWz61KKk
Both moved between the local gate run and CI; CI_GREEN_SKIP=1 because
this commit is the fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015mLnUeoJ3mUwGRTowTUH6d
300M held about two days on kuhy's desktop: one Steam boot wrote 480k
journal lines and the evidence for a curfew failure two weeks earlier was
already gone when it was needed (2026-09-13). The three Arch copies of
the journal tweak (arch_hardware, arch_perf_fixes, arch_perf_report) move
to a 4G cap, and "large"/vacuum now mean past that cap — a 2G journal is
deliberate retention, not a finding. Tests follow.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015mLnUeoJ3mUwGRTowTUH6d
CI has been red since db243fe on exactly these two pins (the freshness
gate; nothing else in that run failed). Nothing in the monorepo imports
gatelock; it is a listed requirement only. Committed with CI_GREEN_SKIP=1
because this commit is the fix the ci-baseline-green hook asks for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015mLnUeoJ3mUwGRTowTUH6d
A push took ~13s. prettier is a pre-push hook precisely because Node's
startup dominates its runtime, and pre-commit chunks a large file list into
several invocations -- a 106-file push paid ten cold Node starts. ci-mirror,
the hook that looks expensive, was already 0.02s once its green-tree cache
hit.
A file whose content, prettier version and prettier config are all unchanged
since its last clean check cannot have become unformatted, so the verdict is
now cached and the file never reaches Node. Only clean verdicts are recorded,
so a formatting regression is always re-checked; the key includes
`prettier --version` and every config file, so an upgrade or a config edit
invalidates the cache wholesale. The systemd-run memory cap is unchanged.
Measured on the same push range: 13.0s -> 3.3s total, of which ~0.6s is
hooks and the rest is git's own network negotiation. On a 40-file list the
runner itself goes 0.652s -> 0.057s, starting no prettier process at all.
endurain/README.md and copilot-instructions.md are reflowed: the ~
reorganisation lengthened paths inside them, which changed prettier's
markdown table alignment.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AwtowrowDXx6ccmWVPkCaF
Also bumps the 16 meta/requirements pins the freshness gate flagged
(anthropic 1.5.0, numpy 2.5.3, pytest-randomly 5, flake8-bugbear 26.9.9,
ruff 0.16.7, pyright 1.1.414, selenium 4.49.0, hypothesis 6.168.0, lxml
6.1.3, pipdeptree 4.2.5, the two types-* stubs) and Flutter 3.47.3 in
meta/.fvmrc. One test still expected the pre-move ~/mcp-servers path.
The pre-commit hook fed .zsh files to shfmt and shellcheck, neither of
which parses zsh (SC2148/SC1071), so no zsh file could be committed since
the hook was tightened on 2026-08-22; those now get `zsh -n` instead. The
hook's own shfmt pass also staged its reformat of 8 shell files.
CI_GREEN_SKIP because main is red on exactly the dependency-freshness
finding this commit fixes, which is the documented use of that flag.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AwtowrowDXx6ccmWVPkCaF
`verify` gated on the two invariants whose breach cut every bridged
container off the internet on 2026-08-29 and again after the 2026-08-30
reboot -- but nothing invoked it, so it reported drift only when a human
thought to ask, and the outage it exists for is exactly the one nobody
thinks to ask about.
nftables-drift.service now runs it after nftables.service, docker.service
and network-online.target at every boot, plus a daily Persistent timer. A
breach is a failed unit in `systemctl --failed`, not a log line: a
self-check that detects a failure has to gate, not warn.
Proven in ~/utils/vmbox across two real reboots -- drifted /etc leaves the
unit failed with the diff and the missing docker0 accept in its journal,
matching /etc leaves it active with no false positive -- then installed and
run green on the host. Not yet observed across a host reboot.
Also bumps the gatelock pin to v0.7.3, which is what turned main's
dependency-freshness job red; CI_GREEN_SKIP is set because this commit is
that fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015d48t4qsQLBJnDwDeo2Nom
wake_alarm._alarm no longer exists: the PC's waking role was deleted in
kuhyx/wake-alarm, and the phone owns the alarm now. This removes ALARM_MODULE,
_run_alarm() and the --with-alarm flag, leaving one leg -- the workout lock.
Landing this in the same change as the deletion is the point. An orchestrator
left pointing at a module that no longer exists is exactly the failure this
repo has already had once: after screen-locker was extracted on 2026-05-28 the
reference here went stale and the workout lock died with ModuleNotFoundError on
every run for a month, while the alarm half kept working and hid it.
morning-routine.service loses --with-alarm and its comments now name what
actually starts it: wake-alarm-trigger.timer, at the phone's synced alarm time,
not the deleted systemd-sleep hook. Environment=DISPLAY=:0 stays -- the comment
credited it to the alarm's Tk window, but workout-locker.service needs it too.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PcdngP688mW6rJye3LRkSN
The forward-chain fix from 334c0cd9 only ever reached the live ruleset by
hand; /etc/nftables.conf kept the bare default-drop chain, so the next
reboot cut every bridged container off again and the Signal bot went deaf
a second time. Nothing compared generated against persisted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The GTK tray icon was the only status surface, and it had been commented out
of ~/.config/i3/config since 2026-08-21. With nothing in the bar, the locker
miscounted the week for a fortnight and the machine never said so.
Shows "<count>/<minimum> · <gaming budget>", read from the local status API
(screen-locker-web.service on 127.0.0.1:8770); clicking opens the full UI.
Three things it deliberately will not do quietly:
- a dead status server renders "status server down" in red, or the last known
value marked (stale) in orange -- never a blank block. That server is what
the gaming budget reads, so its absence costs two hours;
- a disarmed or unscheduled timer overrides the compliance colour entirely
and shows "DISARMED" in red. A locker that cannot fire is the worse
problem and the one that hides;
- clicking only opens the UI. There is no bar action that can clear a lock.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Fcwu78wk17FWMHbbfbyicc
`write_nftables_ruleset` emitted a forward chain with `policy drop` and no
rules at all. That chain runs at the same netfilter hook as Docker's own
DOCKER-USER/DOCKER-FORWARD chains, so it vetoed them: every bridged
container lost outbound networking while the host itself stayed online, so
nothing looked broken.
It cost the Signal bot three silent hours today. signal-cli lost its
connection to Signal at 15:32:50 and looped on reconnect for 2.5h, while
the bot kept answering /healthz with 200 -- `/v1/groups` still served from
local state, so every surface said healthy.
`flush ruleset` also deletes Docker's own tables, and dockerd only rebuilds
them on its next network event, so restore_docker_rules() now restarts
docker right after a successful apply.
Verified in vmbox (Arch guest, docker installed, alpine probe):
BASELINE=reachable -> OLD=blocked -> NEW=reachable, sshd active throughout.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KbjfWdAyHbCubac7wJKJZW
A new app at ~/restaurant-rater rates restaurant dishes, so it needs a
family icon. shield-cutlery is diet-guard's and the accent is shared
fleet-wide, so the new glyph has to be distinguished by silhouette alone: a
filled star above an open bowl -- a dish, and a verdict on it.
The star is filled rather than stroked for the same reason as the nib and
the anvil: at 72px weight its five interior notches close up well above
MIN_NEGATIVE_SPACE and it renders as a pentagon at 48dp.
The bowl went through two passes. The first used shallow walls stopping at
y=736 and read unmistakably as a smiling mouth with the star for a nose --
two curves with a shape above them is a face unless the vessel is clearly
deeper than it is wide at the base. The committed version drops to y=752
between narrower walls, which reads as a bowl.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012YDbT3ip2yeT18W3KCZdLq
Two failure modes compose cannot see, both of which report a healthy
container while it serves 404s:
1. The Caddyfile is a bind mount, so its contents are not in the config
hash compose compares. Editing it changes nothing observable, compose
says Running, and the new directives never load.
2. A bind mount resolves once, at start. Replacing dist/ rather than
emptying it leaves the container on the old unlinked inode: /srv reads
empty and every request 404s.
Hit both at once on a real deploy today -- the container was Up 6 hours,
Caddy logs clean, /srv empty, curl 404. Reproduced by rm -rf dist (inode
84695030 -> 84695031) and fixed by the patched function in one run.
wait_for_website also stops using curl -sf, which collapses every non-2xx
into the same "did not become reachable". It now reports the status it saw,
because unreachable and answering-404 are different bugs and the message
named only the first.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A7BaCSV8FEtcNgVDVsRAJj
Upstream released it between two CI runs today, which is enough to turn the
freshness gate red on a commit that touched no Python -- the documented and
expected behaviour of a gate that checks "newest stable" rather than "newest
stable as of whenever someone last looked".
The pre-commit pylint hook takes its version from the mirror rev, not from
this file, so nothing about which checks run changes here.
Also refreshes the isort note: 4.0.7 declared `isort<9`, 4.0.8 declares
`isort!=5.13,<10,>=5`, so the comment's specific bound was already stale.
Reworded to state the constraint's purpose rather than a version that moves.
Fixes the red main that ci-baseline-green is blocking commits on.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A7BaCSV8FEtcNgVDVsRAJj
The personal-website repo now prerenders each post to
dist/blog/<slug>/index.html. Caddy try_files does not resolve a directory
to its index on its own, so with the single-term directive every post URL
fell through to the SPA shell: right content once React booted, wrong
<title>, and no link preview. Verified against caddy:2.8 in Docker both
ways before and after.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A7BaCSV8FEtcNgVDVsRAJj
The guard prints `CI_GREEN_SKIP=1 git commit ...` and stops there, but it runs
at two stages: at commit, and again inside the pre-push CI mirror's
clean-worktree `pre-commit` run, which sees the same still-red baseline. So the
documented escape gets you past the commit and into an identical wall at push,
with no hint and nothing in the output connecting the two.
Hit today landing the pre-commit cache-key fix: the commit went through, the
push did not. The variable is read from the environment and `ci_mirror.sh` does
not scrub it, so `CI_GREEN_SKIP=1 git push` carries through to the nested run --
which is now what the message says.
Verified both paths: with the variable set the script warns and exits 0, and
against today's green main it exits 0 silently. shellcheck clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SfnJmxxsZ534AKk16rC2Gr
The key said `py3.12` as a literal while the job installs `3.x`, which is
3.14.7 today. A `language: python` hook env is built by whichever Python
pre-commit ran under, so the cache kept handing back envs built on 3.12.14 and
the hooks kept running there -- invisibly, right up to the first syntax 3.12
cannot parse.
PEP 758 was that syntax. The previous commit's `except A, B:` reached pylint
and the debug-statement hook on cached 3.12 interpreters and came back as
"multiple exception types must be parenthesized", a failure with no local
reproduction: developer caches were built on 3.14, where the same pylint 3.3.2
and the same pylint 4.0.7 both parse the file fine. The syntax was never the
defect; the stale interpreter was.
`steps.setup-python.outputs.python-version` cannot drift from what was
installed, so this cannot recur on the next release either.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SfnJmxxsZ534AKk16rC2Gr
Already red before today's Python work touched the repo; the commit hook had no
reason to look at the requirements files, so only a whole-tree run saw it.
Bumped across every repo pinning 7.15.4 in the same pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SfnJmxxsZ534AKk16rC2Gr
`requires-python` and the CI pin already said 3.14, but ruff, mypy and pylint
were still grading `meta/` against 3.10 -- three tools judging an interpreter
nothing runs any more, and ruff withholding the pyupgrade fixes that only
become safe at the newer target. `pin_ci_python.py` rewrites these three knobs.
What the stale target was hiding: `datetime.timezone.utc` where `datetime.UTC`
is now correct (UP017), and ruff's stable formatter applying PEP 758 --
`except (A, B):` becomes `except A, B:`. That makes the source 3.14-only at the
syntax level, which is what `requires-python = ">=3.14"` already promised; kuhy
chose it deliberately over holding ruff a release back (2026-08-29).
Also drops a comment in `wsg_grabber/models.py` that the bump makes false: it
explained why the file used ``str, Enum`` instead of ``StrEnum`` under mypy at
3.10, sitting directly above a ``class Verdict(StrEnum)`` that had already
stopped obeying it.
ruff, ruff-format, mypy --strict and pylint all clean at the new target.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SfnJmxxsZ534AKk16rC2Gr
`{"flutter": "stable"}` names a moving target: two machines on the same commit
can build with different SDKs, and a freshness check has nothing to compare it
against. This is the only .fvmrc in the fleet, and it was the one file the new
dependency-freshness gate could not evaluate.
Purely declarative today -- nothing in any repo invokes `fvm flutter`, and fvm
has no SDKs cached, so the build path is unchanged and still resolves the
pacman-installed SDK through ~/.local/bin/flutter.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CRykabMwPgDzFHcARoqpaK
Counting transcript records reports 1.00 tool calls per message no matter
what the model did: the harness writes one JSONL record per tool_use block,
all sharing message.id. A hand-rolled per-record count read the real 1.9%
batching rate as 0.0%, which nearly shipped as a headline finding.
Group by message.id instead, and report it, so the number comes from a
tested code path with the window handled by find_transcripts rather than an
ad-hoc timestamp. Three snapshot fields give next week's run a baseline.
Axes was at the 12-attribute cap, so the counters live in their own Batching
dataclass with the two ratios as properties.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NiSSDu4gKbp9GGxTTtXvYW
