Audit 2026-10-04: fix fork bugs and red tests - #32
Merged
Merged
Conversation
Upstream pingdotgg#8748 moved the known Windows CLI directories from the front of the resolved PATH to the end, and the merge updated upstream's tests but not the fork's persistent-PATH test, so packages/shared failed on main. The implementation is correct; the expected order now matches it. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The Codex voice bridge treated any ask_selected_agent call within 30 s of an answer as a GPT-Live repeat ping and replied with the previous answer, so a different follow-up question was never asked. Reuse the answer only when the request matches the one it answered, and key the speak-once dedupe on the request too, so two questions with the same answer are both spoken. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
self_hosted_push_devices.expo_push_token is UNIQUE, but the register upsert only resolved a device_id conflict. A device that came back with a new device id and the same push token failed the insert (500 from /api/notifications/device) while the stale row kept receiving its pushes. Drop any other row holding the token in the same transaction before upserting, so the newest registration wins. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A merge kept the fork's branded label but also its hard-coded deviceType "mobile", dropping upstream's tablet detection, so every tablet paired as a phone in the authorized-clients list. Restore the Device.deviceType mapping. connection.test.ts now mocks expo-constants (pulled in by the branding module), which is what kept this regression from being caught. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…eme picker BUILT_IN_THEMES includes MT Code's palette, which by design ships to web and desktop only, so the mobile picker listed an "MT Code" theme that normalizeMobileThemeId discards on save. Filter it out. mobileTheme.ts also imported the branding module (expo-constants) for the default palette's label, which made 12 mobile test files fail to load under Vitest. Mobile went back to T3's identity in ee2b9ff, so the label is the literal again, and the two remaining tests that reach branding mock expo-constants like the rest of the suite. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With CLAUDE_CONFIG_DIR set, Claude Code keeps .claude.json in that directory; the file in the home directory belongs to the default login. The tier reader fell back to it anyway, so a configured account could be labelled with another account's Max 5x/20x tier. Read only the file the CLI would use. The checkClaudeProviderStatus tests read the host's real ~/.claude.json through that path and failed on any machine signed in to a Max 20x account; they now point CLAUDE_CONFIG_DIR at a missing directory. The default provider list no longer expects a devin instance, matching upstream since Devin moved to the ACP registry. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ver ends servicePreflight.test.ts still expected "T3 Code" in the launcher message, which the fork builds from resolveAppDisplayName(). The stalled usage probe test waited on the probe's real 4 s usage deadline under a TestClock nobody advanced, so it timed out after 120 s; run the probe on the live clock. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The Clerk deep-link handler loaded any <scheme>://<host>/... URL into the main window, but the protocol handler only serves host "app". In builds where thread links share the scheme (t3code://threads/...), and for any other host a web page links to, the window went blank. isDesktopProtocolUrl now requires host "app"; thread links stay with DesktopDeepLink. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The launcher icon test expected the stock production icon, but personal builds deliberately use the Munim or Nightly icon. The stalled-capture preview test expected upstream's three quick capture retries, while the fork bounds each grab at AUTOMATION_CAPTURE_TIMEOUT_MS and reattaches once, so the test never finished. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ExternalTerminalSetting read window.desktopBridge unguarded, which threw in the IntegrationsSettings test renderer and took the whole panel down; guard it like the other desktop-only settings. Also drop a merge-era override in threadActionMenu.logic.test.ts that hid the auto-settle item the test then expected. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
electron-updater emits update-available from inside checkForUpdates, while the check still holds the update action reservation. The fork's auto-download ran straight from that event, so tryStartUpdateAction refused it on every check and updates sat at "available" until the user clicked. Wait for the check to release its reservation (bounded, like the prepared-install wait) before downloading. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…the old one swap_install runs as `if ! swap_install`, so set -e is off inside it. If moving the installed app aside failed (or ditto failed, e.g. a full disk), the script carried on, moved the staging app *into* the live bundle and reported the install as done. Each step now returns on failure, restores the previous app when the final move fails, and the caller reports the real reason. Verified by extracting swap_install with a simulated mv failure: before, it returned success with staging.app nested in the live app; after, it fails with "could not move the installed app aside" and leaves the app untouched. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The refresh scripts stamp release versions into four package.json files and restore them only after a successful build. Under set -e a failed build exited first, leaving the checkout dirty, and every later run then skipped with "checkout is dirty" and exit 0, so the fleet silently stopped refreshing. Restore them from an EXIT trap instead. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…tion The link went to the inherited docs/user/install.md, which installs upstream T3 Code from pingdotgg releases and t3.codes/install.sh. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
github-actions
Bot
added
vouch:trusted
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Audit of the fork's own changes (
git diff $(git merge-base origin/main upstream/main) origin/main, 1,452 fork commits). Each fix is a small, separate commit. Everything was verified on a scratch clone on the Mac: fullvp run -r typecheckplus every package's test suite.Bugs fixed
ask_selected_agentcall within 30 s of an answer was treated as a GPT-Live repeat, so a different follow-up got the previous answer and never reached the agent. The answer is now reused only for the same request.CodexVoiceSession.test.ts: fails before the fix, passes after.expo_push_token UNIQUE(500 from/api/notifications/device), and the stale row kept getting the pushes. The old row is now dropped in the same transaction.UNIQUE constraint failedbefore, a single row after. Typecheck.update-availablefires insidecheckForUpdateswhile the check still holds the update reservation, so every download was refused and updates sat at "available". It now waits for the check to finish first.DesktopUpdates.test.ts: fails before (available), passes after (downloading). All update tests pass.<scheme>://<host>/…URL into the main window. Thread links (t3code://threads/…in non-MT builds) and arbitrary hosts blanked it. Only hostappis loaded now.desktopProtocolUrl.test.ts.deviceType: "mobile"and dropped upstream's tablet detection.identifies native tablets separately from phonesnow runs and passes.normalizeMobileThemeIdthrows away. Separately,mobileTheme.tspulling inexpo-constantsstopped 12 mobile test files from loading at all.~/.claude.jsoneven withCLAUDE_CONFIG_DIRset, so a configured account could show another login's Max tier. The tests also read the host's real file.ProviderRegistry.test.tspasses on a Max 20x machine. Before, it failed 3 tests there.swap_installruns with errexit off. A failed "move aside" moved the new app inside the live bundle and still reported success.mvfailure. Before: success, withstaging.appnested inside the app. After: clean failure.bash -npasses, and shellcheck shows no new warnings.package.jsonversions dirty. Every later refresh then skipped with exit 0. The versions are now restored from an EXIT trap.bash -nand shellcheck. Not executed: these scripts install the app.ExternalTerminalSettingreadwindow.desktopBridgeunguarded.IntegrationsSettings.test.tsxpasses.Stale tests fixed (main was red)
These are test-only changes. Each one aligns a test with intentional fork behaviour or a later upstream change; none weakens what the test checks.
packages/sharedshell PATH order: upstream fix(shared): preserve Windows shell PATH priority pingdotgg/t3code#8748 moved the known CLI dirs to the end of the resolved PATH.devininstance, matching upstream since Devin moved to the ACP registry.servicePreflightmessage: usesresolveAppDisplayName().TestClock.withLive.threadActionMenutest: a merge-era override hid the item the test expects; it is removed.Found but not fixed (for the owner)
scripts/personal-codesign-mac-dmg.shre-signs the public Mac release with osx-sign's default entitlements. Those lackcom.apple.security.automation.apple-events, so in the hardened, notarized build Apple Events fail with -1743. Not changed because the fix can't be verified without signing and notarizing.origin/mainwhile Mac builds localHEAD, so one tag can ship different commits.personal-release-fleet.shalso writeslast-built-shaeven when the publish failed.$ErrorActionPreference="Stop"inpersonal-refresh-win.ps1, a launcher's stderr aborts the whole refresh.infra/updates/src/worker.jsaccepts unauthenticated counter writes with an unboundedversion./statsis public and unthrottled.CLIENT_SALTfalls back to a public value.apps/server/src/cli/desktopLaunch.tsdoesn't recognise MT Code AppImage.desktopentries.scheduled_for,recurrence_json) have no readers or writers left after the orchestrator-V2 merge.docs/internals/providers.mdlinks to files that no longer exist.invocation.test.ts,ServerEnvironment.test.tsandAntigravityAdapterV2.test.tsfail on macOS (/varvs/private/varrealpath), and two mobile Swift script tests fail with Swift 6.4. All of these are unchanged upstream code.serverSettings.test.ts"reloads when a dangling settings link gets its destination" failed once under full-suite load and passed 3 out of 3 runs in isolation, so it looks flaky.🤖 Generated with Claude Code