[adampeterhiggins]

What’s in this sync

Adds 431 upstream commits since the parent sync.

• New features: UI language settings and translations for Spanish, Chinese, French, German, Japanese, Korean, Russian and Brazilian Portuguese; mobile language selection; reply from assistant hover actions; resolved review conversations fold into avatars; quoted messages display as inset cards; Windows background service support. Add-project lists default to GitHub until other hosts are enabled.
• Fixes: more reliable Stop and delegation completion, scheduled-task missed-slot reporting, worktree creation and cleanup with large output, composer chip restoration, streaming scroll behavior, mobile settings and review navigation, and desktop update recovery.
• Provider and internal changes: Cursor streaming and steering improvements, Claude delegate process cleanup and custom model handling, Codex policy errors and delegate lifecycle fixes, OpenCode service restart on update, and lower-cost thread/state processing.

Discord announcements were not used: the channel returned no matching announcement crossposts for this range.

Parent #572 has merged. This PR now targets main and adds only the newer upstream commits.

Base: ae9bd790f5f277097dd72f3998c4825829a40852.
Pinned upstream: ee90bdc9405d8d4b8ff78ce2133ccb04844298ab.
Range: ae9bd790f5f277097dd72f3998c4825829a40852..ee90bdc9405d8d4b8ff78ce2133ccb04844298ab.

Other notable changes: Windows service support (pingdotgg#34524), provider health checks (pingdotgg#34885), large worktree Git output (pingdotgg#34889), Codex child cleanup (pingdotgg#34829), missed scheduled runs (pingdotgg#34815), desktop update recovery (pingdotgg#34932), and singular CSV/text preview labels (pingdotgg#34992).

Full upstream comparison.

Conflict resolutions and verification

Preserves fork context integrations, provider/account selection, quick-message and split views, citation comments, notification rules, missed-run catch-up, and worktree setup naming while adopting upstream guards and provider changes.

Regression repairs retain linked senders and missed-slot errors; complete workspace handoffs through checkpoint capture before starting the continuation; preserve the destination divider and user Stop precedence; canonicalize macOS symlink paths; and align composer state with the selected provider. Stale TCP, recording, remote-delegation, scheduling and browser fixtures were repaired without removing their behavior assertions. The prompt budget records the actual merged text (+31 bytes for Claude and Codex); no instructions were cut to satisfy the measurement.

Final automated verification on dc1095749980c6e375402a7527fb94e516759871:

• Non-browser regression selection: 548 files, 13,727 tests passed, 26 existing skips, across 28 successful batches.
• Previously omitted browser selection: 117 files, 913 tests passed, no skips. Full reruns of the two repaired batches replace their initial failing results.
• Handoff lifecycle: 146 focused tests pass, including persisted checkpoint, continuation, destination notice, and Stop race assertions.
• Server and web typechecks, targeted format/lint, and commit hooks pass.
• Final check:changed: format, lint, knip, generated GraphQL, dependency-patch validation, all 41 affected package typechecks and node tests pass. Its single 492-file Vitest invocation exceeded the standard 900-second limit. The separate 665-file batched run above covers that selection and the previously excluded browser files with zero failing cases; the standard command itself is not claimed as passing.

Runtime: the isolated Electron app streamed the free Replay showcase, accepted approval through real Electron input, and rendered its final markdown/code/table. The web client opened the same completed thread and Settings. The iOS dev client connected and displayed the completed thread, code and table without the Expo overlay. Further iOS language-navigation checks hit a CoreSimulator/device-runner deadline after that successful pass; Android is unavailable because this machine has no Android SDK.

Limits: the native-Claude compaction case passes the integrated browser fixture, but the ACP-only scenarios harness cannot create that native session/context state for a separate scenarios:prove run. Paid MCP evals remain unavailable without their API credential. No merge or release was performed.

Merge with Create a merge commit. It keeps upstream history and the merge base for the next sync.

Models/harnesses: gpt-6.1-sol in Codex; earlier follow-ups used grok-4.7 and claude-opus-5-5 in Cursor, through T3 Code.