[gabynevada]

The fork is 42 commits behind pingdotgg/t3code:main. This merges upstream through 1d78b33203 into atli. It keeps upstream ancestry and the fork's customizations.

Upstream highlights:

• Scheduled tasks can be moved to another thread (web and mobile). Mobile task prompts no longer sit under the microphone.
• Server performance: Codex tool calls no longer wait on reasoning summaries, and worktree branch allocation is batched. Worktree cleanup skips unneeded scans, sidebar branch labels skip the full git status, and handoff selection skips oversized history.
• Server fixes: idle parents wake for late delegated results, and mise-installed providers update through mise. macOS Bash terminals start as login shells, Antigravity browser preflights can run longer, and rejected peer certificates are named when linking. Linking tools refresh in existing agent sessions. HTML tools take a file path. Saved preview screenshots keep full resolution.
• Thread links: unloaded links show their title and project icon (new getThreadLinkMetadata RPC), and angle-bracketed destinations parse correctly.
• Desktop: cursor:// and vscode:// links open from chat. Linux x64 cold backend startup is faster, Linux ARM64 gets server bytecode, and macOS declares Calendar access.
• Mobile: the Android terminal forwards arrow keys, Tab and Escape from a hardware keyboard. PDFs can be shared from the Android viewer, side chats stay readable, and several centering and layout fixes landed.
• Web: bold weight is back in rendered Markdown. Also fixed: file chip tooltips, the wide-table scrollbar, combobox search fields, the sidebar width, and sidebar search results can be dragged onto the composer.
• Antigravity accepts the agy CLI as a custom binaryPath, and Muse no longer wedges a thread when /compact is declined.

Conflict resolutions:

• Both sides added an orchestration RPC at the same spot. The fork added listThreadGroupingParents and getThreadGroupingDescendants, and upstream added getThreadLinkMetadata. All of them are kept across contracts, RPC auth scopes, instrumentation, ws.ts, Orchestrator, ProjectionStore, ThreadManagementService, and the two test mocks.
• ProviderSessionManager: the fork's refreshPullRequestWatch and upstream's readLinkedEnvironments are independent helpers, so both are kept.
• agentContextBudget.json: the values were re-measured on the merged text (5465 / 7195 / 49148 bytes). All three are lower than the fork's current budget.

Merge fallout fixed in the merge commit:

• ScheduledTaskService.test.ts auto-merged with a duplicate ScheduledTaskUpsertInput import. That broke lint, typecheck, and the test transform.
• Upstream's move-task change now validates the destination thread through getThreadShell on upsert. The fork-only "bound tasks follow the thread" test now mocks it.

Merge with a merge commit (not squash) to preserve upstream ancestry for future syncs.

Validation:

• vp run check:changed origin/atli: format, lint, both Knip passes, GraphQL, 40 package typechecks, and React compiler checks pass. Server tests pass, including all 13 ScheduledTaskService cases after the fix.
• Under the full parallel check:changed run on this busy machine, 8 web browser test files intermittently report 0 test with a vitest "error when mocking a module". They passed in the first full run, and all 8 pass together (49 tests) when apps/web runs on its own. I treated this as load flakiness, not a merge regression; CI will confirm.
• No manual browser or device smoke pass and no native build. CI covers the full suites and build jobs.

Model: Claude Opus 5.5, Claude Code harness in T3 Code.

🤖 Generated with Claude Code