Skip to content

Project audit fixes, and Live Radar on the Think model - #474

Merged
stevyhacker merged 6 commits into
masterfrom
claude/project-audit-cleanup-322126
Nov 10, 2026
Merged

stevyhacker merged 6 commits into
masterfrom
claude/project-audit-cleanup-322126

Conversation

@stevyhacker

@stevyhacker stevyhacker commented Nov 10, 2026 •

Copy link
Copy Markdown
Owner

Summary

Two things, in five commits (one per area), and a sixth that repairs the hosted film test:

  1. Fixes from a whole-project audit. Seventeen area reviews returned 283 findings. Each was read again in the code before anything changed. 125 are fixed here, 24 are fixed in part, 1 turned out not to be a defect, and 133 are left because they need a decision or could not be changed safely or measured.
  2. Live Radar reads with the Think model. Live Radar, its screen answers and its live answers used Meta's Model API, a hackathon requirement. They now run on the Think model chosen in Settings → Models, through ThinkExecution like notes and Ask. Follow-through and fast live text still use Meta's models.

What users would notice most:

  • Mac: a model's tool call could stop the app (get_meeting with a huge from); reprocessing a meeting whose notes repeat an action could stop the app; live answers' code tools read hidden, vendored and secret-looking files; Hand off packages outlived their 7 days; the overnight review read the sample meeting; voice samples of merged meetings never expired; ask waited a minute before reporting a closed app; autocomplete skipped warming up the next field after focus left a suggestion.
  • Windows and Linux: one unreadable setting reset all settings; a malformed JPEG stopped file indexing; a hung audio system hung the engine; a cancelled dictation blocked day digests; a model was reported ready while still downloading; the recording in progress could be deleted; clicking a notification did nothing on Linux; dictated text went into clipboard history.
  • Phones: the Android People list crashed on a date without a time zone; a question queued on Android never showed its answer; iOS pairing from pasted text failed when the text held more than the link.
  • Site and CI: wrong claims corrected, 620 lines of unused styles removed, and the macOS 26 UI workflow can be started by hand again.

Both command-line tools now print the error code before an access refusal ([access_disabled], [files_access_disabled], …), which the agent skill already tells agents to look for, and both MCP servers refuse an include value they do not know.

Validation

Mac, on the final tree:

xcodegen generate
xcodebuild -project LokalBot.xcodeproj -scheme LokalBot -configuration Debug \
  -destination 'platform=macOS' CODE_SIGNING_ALLOWED=NO -only-testing:LokalBotTests build-for-testing
# ** TEST BUILD SUCCEEDED **
xcodebuild test-without-building -xctestrun … -destination 'platform=macOS,arch=arm64'
# Executed 4505 tests, with 58 tests skipped and 0 failures (0 unexpected)   (master: 4,482)
CODE_SIGNING_ALLOWED=NO bash Scripts/ui-tests.sh --build-only
# ** TEST BUILD SUCCEEDED **   (UI tests were not run)
swiftlint lint --strict --quiet
# exit 0   (master: exit 2, the Lint check has been red since PR 427)
  • For 30 of the fixes (8 on the Mac and 15 on the port in the second pass, 7 in the first), the fix was undone, the test or check was seen to fail or stop the test host, and the fix was put back.
  • The first commit also builds on its own, and the 52 tests in the classes whose files are split between the first two commits pass there.

Windows and Linux, on the Ubuntu box, whole workspace, with the regenerated catalogs:

cargo fmt --all -- --check                                        # exit 0
cargo clippy --workspace --all-targets --locked -- -D warnings    # exit 0
cargo test --workspace --locked                                   # 2,347 passed, 0 failed, 28 ignored
python3 -m unittest discover -s scripts/tests                     # Ran 8 tests, OK
python3 scripts/sync-localizations.py --check                     # exit 0
dbus-run-session -- cargo test -p lokalbot-platform --test live_dbus -- --ignored --test-threads=1 notification
# 3 passed

Phones:

Scripts/build-sync-android.sh                                     # on the Ubuntu box
./gradlew :app:assembleDebug :app:testDebugUnitTest               # BUILD SUCCESSFUL, 126 tests, 0 failures
./gradlew :app:verifyRoborazziDebug                               # BUILD SUCCESSFUL
./gradlew :app:lintDebug                                          # BUILD SUCCESSFUL
python3 -m unittest discover -s Companion/e2e                     # Ran 22 tests, OK
swift test --package-path Companion/ios/LokalBotCompanionKit      # 66 tests pass

Site and scripts:

python3 Scripts/render_web.py --check        # All 39 generated web files are up to date.
python3 -m unittest discover -s Scripts/tests   # Ran 193 tests, OK
(cd Distribution/npm && node --test)         # 50 pass
  • Every page (44) was rendered in headless Chromium with the stylesheet and script from before and after the clean-up, at two widths and both appearances: computed styles and boxes are identical in all 176 renders, and no page gained a console error.

UI suite, hosted, on caab609: every shard passed except UI functional-1, where HackathonFilmUITests.testFilmLiveRadarAndFollowThrough failed at its 4.1% in September check. Its settings steps under settings.liveRadar.* passed.

  • The same check failed in PR 427's own UI run and in two other branches' runs on 2026-11-09, so master has the failure too. The film workflow always ran with live replies; the scripted replies first ran in the required suite.
  • Cause: the scripted screen answer matched its words against the request's first text part, which begins with the meeting's title. In "Launch sync" every question matched "launch", so the Team plan question was answered "Monday, November 3", and the test took that first screen answer for the conversion one.
  • Fix (b2abe13): the rule reads only the question. Replayed outside the UI with the app's own term, crop and quote code on the two fixture captures: before, the Team plan question gets "Monday, November 3"; after, it gets no answer, the conversion question gets "4.1% in September" and the launch question "Monday, November 3".
  • CODE_SIGNING_ALLOWED=NO bash Scripts/ui-tests.sh --build-only succeeds with the fix. The test itself runs only in CI.

Not verified:

  • The film test after its fix. Its checks after 4.1% in September have never run with scripted replies, so this push is their first run.
  • Live Radar in a real meeting on a local Think model or a Think server. The pass, live-answer and screen-answer paths are covered by unit tests with a scripted engine. Speed on a small local model was not measured.
  • Windows. Nothing was compiled for Windows; no code inside #[cfg(windows)] was edited.
  • The port's desktop-app fixes, the clipboard change and the recorder start timeout were compiled, not exercised in a running app. install.sh --uninstall was syntax-checked only.
  • The iOS app sources were type-checked, not built or run. The Raycast change was not type-checked.

Changed test expectations

Mac:

  • MeetingRadarTests/testRadarNeedsApprovalBeforeAnythingIsSent → testRadarIsOffUntilItIsTurnedOnInSettings: the radar was unavailable until api.meta.ai was approved and a key saved → it is unavailable until "Offer Live Radar during meetings" is on.
    Why the old expectation no longer holds: Live Radar no longer sends to Meta, so Meta's approval is not what gates it. A remote Think server still needs its own approval, which the Think engine checks.
  • RadarScreenAnswersTests/testRadarAnswersACaughtQuestionFromTheScreenAndSavesTheCrop: the model was sent the crop as a picture and the answer had a cost → it is sent the crop's text, no picture, and the cost is 0.
    Why: a Think model may read text only. The picture path is kept for the hosted film test and has its own test, testAModelThatReadsPicturesIsSentTheCropAndItsCostCounts, with the old assertions.
  • MeetingRadarTests/testPassesSendOnlyNewLines…: same assertions. It now names Muse Spark as the model so its cost assertion still uses Meta's prices.
  • DailyMemoryExportServiceTests, helper writtenURL (9 tests): an export that was not written skipped the test → it fails it. The old helper hid a regression as a skip.
  • MeetingIntegrityTests: six calls drop hostReconnectGrace: 15. The parameter was ignored and is removed. No assertion changed.
  • RedesignContractTests/testLazilyFilledScreenGroupsMatchFillingEveryHit: same assertions; the random input comes from a seed that each failure names.
  • AskLibraryEngineTests, helper makeEngine: says the app is running, so the wake tests still reach the wake now that a closed app is reported at once.
  • HackathonFilmUITests/testFilmLiveRadarAndFollowThrough: the checks for "Send the revised pricing deck" and "AppSettings.swift" looked for a static text → they look for the text in any element's label.
    Why: the commitment's card is in a lazy grid below the cards that need attention, which may not be built while out of view (the floating panel lists it as a row either way), and a code source is a button. No run has reached either check, so this is a precaution, not a failure that was seen.
  • LokalBotTests/Fixtures/upgrade/{0.6.2,0.7.2,0.8.5,0.9.0,0.9.1,0.9.2,0.9.3,0.10.10}/library/lokalbotv3.sqlite-wal and -shm removed. Every -wal was 0 bytes. UpgradeMigrationTests pass.

Windows and Linux:

  • lokalbot-cli/tests/cli.rs, brief_names_cli_commands_where_get_briefing_names_tools: "Commands: search, get, list, actions, and people." → "Commands: search, get, list, actions, people, and ask while the app runs."
    Why the old expectation was wrong: ask is a command on the port too, and the Mac's brief already listed it.
  • lokalbot-platform/tests/live_dbus.rs, notification_actions_reach_the_callback (ignored by default): the actions offered now end with "default", "LokalBot".
    Why: a notification server reports a click on the body only for a notification that registered the default action, so the callbacks for "open the update" and "review this request" never ran.

Android:

  • pairing/PairingLinkTest.kt: the expired case was now - 1 → now - 121. ui/ViewModelTest.kt: the expired link was now - 10 → now - 600.
    Why the old expectation was wrong: the phone refused a code the moment it expired on the phone's clock, so the 120 s the computer allows for a fast clock never applied. now - 1 and now - 120 are asserted valid in a new test.

Linked issues

None closed. Issue 458 (Parakeet downloads on Windows and Linux) is not fixed here: upstream re-exported both models, and moving the pin ships files nobody here has run.

Risk / rollout notes

  • New setting liveRadarEnabled, off by default. Saved settings without the key decode it as on when https://api.meta.ai is in the approved origins, so people who used Live Radar keep it. For them Live Radar now reads with their Think model: on this Mac if Think is local, or at the Think server they approved. PRIVACY.md says so.
  • Removed settings museRadarModel and liveAnswersModel. Saved values are ignored. The Windows and Linux settings file keeps unknown Mac keys as before.
  • Screen answers send text, not pictures, on the Think path.
  • Live Radar strings changed in all 24 languages (9 new, 3 derived, 9 removed), translated by a model and checked for format specifiers, not by native speakers. The Windows and Linux catalogs are regenerated from them.
  • Command-line output changed: an access refusal now starts with its code, and get_meeting with an unknown include value is an error. An agent that matched the old bare message would need the code instead; the shipped skill already names the codes.
  • Windows and Linux notifications with a callback now register a default action labelled "LokalBot". Most desktops never show the label; one that lists every action would show it.
  • Editing build.yml and ui-tests.yml empties the compiler cache once, so the next Build and UI runs compile from scratch.
  • Site pages' dates moved to 2026-11-10 where their text changed.
  • Two unused view files, unused declarations, 620 lines of unused styles and six unreferenced images are removed. Nothing referenced them.

Defects:
- get_meeting's "from" time could overflow and stop the app.
- Live answers' code tools read hidden, vendored and secret-looking files
  that git tracks.
- Hand off packages were deleted only when another handoff happened.
- The overnight review read the sample meeting.
- "Go to Today" could stay on the Actions review.
- Voice samples of meetings folded into a merge never expired, and the
  hourly pass rewrote every expired journal again.
- Reprocessing a meeting whose notes repeat an action stopped the app.
- Focus leaving a shown suggestion left autocomplete "ready", which skipped
  the next field's warm-up, and a focus change replaced a recording's pause
  label.
- ask, reply and ask_library waited a minute before reporting a closed app.
- install-skill could not find its bundle when run by bare name from PATH.
- A refused streamed request lost the provider's reason.
- The audio monitor took playback from the Dev and UI Test Host builds for
  a meeting candidate.
- "Menu bar only" could read the old setting; a file-search window kept its
  player after closing.

The command-line tool prints its code before an access refusal
([access_disabled], [files_access_disabled], ...), and get_meeting refuses
an include value it does not know, as on Windows and Linux.

Also: work done once instead of per call (clock stamps, citation patterns,
the link detector, list flags, saved-moment notes, the launch relink
query), unused code and two unused views removed, stale comments corrected,
the Lint failure from the hackathon merge fixed, a test helper that skipped
where it should fail, and a test whose random input had no seed.
Live Radar, its screen answers and its live answers used Meta's Model API,
a hackathon requirement. They now run on the Think model from Settings →
Models, through ThinkExecution like notes and Ask, so they follow its
model, server and consent.

- ThinkChat turns the radar's structured requests and the live answers'
  tool loop into TextEngine calls.
- Screen answers give the model each crop's text, not the picture.
- "Offer Live Radar during meetings" (Settings → Meetings → Live Radar)
  turns the feature on. It is off for new installs; settings that had
  approved api.meta.ai keep it on.
- The radar and live-answer model pickers are gone. Meta Muse keeps the
  approval, the key, fast live text and Follow-through's model.
- Cost is shown only where a meeting has one.
- The UI-test host keeps Meta's request format for the hosted film test.

Follow-through and fast live text still use Meta's models.
- One unreadable setting no longer resets every setting.
- A malformed JPEG no longer stops file indexing, and an impossible
  recorded length no longer panics.
- A hung audio system at recording start no longer hangs the engine.
- A cancelled dictation no longer blocks day digests and overnight reviews.
- A model counts as ready only once its download has finished, deleting a
  model removes its digest file, and a resumed download checks free space.
- An approved plain-HTTP autocomplete server on the private network works,
  and a server that does not stream is asked once.
- The recording in progress cannot be deleted.
- A click on a notification's body reaches its action on Linux.
- Pasted dictation stays out of clipboard history.
- A phone connection that trickles bytes ends at the handshake deadline.
- In-app updates refresh the KDE input-method entry, and install.sh
  --uninstall removes the agent skill and the login entry.
- The command-line tool prints its code before an access refusal, and
  get_meeting refuses an include value it does not know.

Also: older library reads no longer replace newer ones, recording a
shortcut pauses the global shortcuts, the recording-health file is saved
every 30 s again, and unused code, icons and two dependencies are removed.
- Android: the People list no longer crashes on a date without a time zone.
- Android: a question queued for later shows its answer.
- Android: search results mark the matched words instead of printing « ».
- iOS: pairing from pasted text works when the text holds more than the
  link.
- iOS: App Lock covers the iPad's meeting window.
- Both: a pairing code is accepted for the 120 s after expiry that the
  computer allows.
- Unused code and an unused dependency removed; protocol documents
  corrected.
- Site: corrected claims (file import, model sizes, MCP tools on Windows
  and Linux, Granite Speech 5, how agent access is turned on), footer
  links, and 620 lines of unused styles and 86 of unused script removed.
  Every page renders the same before and after.
- The local preview server serves pages as production does.
- Documents: settings locations, the port's external-agent switches, model
  notices, the Chinese README's facts about the port, release steps.
- CI: starting the macOS 26 UI run by hand no longer fails at the build
  job, the scheduled UI run and a release dispatch no longer cancel each
  other, and the Linux/Windows workflow runs when files its tests read
  change.
- Scripts create the vendor folders XcodeGen needs before a first build.
- Six unreferenced images removed; the MCP bundle names www.lokalbot.com.
@cloudflare-workers-and-pages

cloudflare-workers-and-pages Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

Deploying with  Cloudflare Workers  Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status Name Latest Commit Preview URL Updated (UTC)
✅ Deployment successful!
View logs
lokalbot b2abe13 Commit Preview URL

Branch Preview URL
Nov 10 2026, 09:33 AM

@stevyhacker
stevyhacker marked this pull request as ready for review November 10, 2026 08:35
The hosted film test's scripted screen answer matched its words against
the first text part of the request, which begins with the meeting's
title. In "Launch sync" every question matched "launch", so the Team
plan question was answered "Monday, November 3" from the launch plan.
The test takes the first screen answer for the conversion one, and
failed on "4.1% in September". The rule now reads only the question.

The film workflow always ran with live replies, so the scripted path
first ran in the required UI suite for PR 427. It failed there and in
each of the three runs of that shard since.

Two later checks in the test, which no run has reached, now look for
their text in any element's label: the commitment's card sits in a lazy
grid that may not be built while out of view, and a code source is a
button.
@stevyhacker
stevyhacker merged commit 93389c7 into master Nov 10, 2026
40 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant