[ScriptedAlchemy]

Why

The fork was nine commits behind pingdotgg/t3code/main. This brings it through upstream commit 54c5f25249961a1aa0b61932329092eadbd9d6fc while preserving the fork's additions.

This is an upstream merge, preserving the fork's 389 existing commits and their SHAs. Merge this PR with Create a merge commit so upstream remains in the ancestry for future synchronizations.

Changes and conflict resolution

• Include upstream v0.1.6, lazy thread-family loading, model-picker focus/positioning fixes, Antigravity quota fixes, OpenCode's Big Pickle text-generation fallback, and expo-device-hub 0.19.1.
• Resolve ProviderModelPicker.tsx by retaining account-member labels and full pool/member disclosure alongside upstream's picker behavior. Standard composer labels keep upstream's plain rendering; pooled accounts retain their tooltip, which closes while the picker is open. Expose full account identity to assistive technology.
• Resolve build-desktop-artifact.test.ts using the exact packaged Windows executable while preserving fork packaging coverage.
• Preserve mobile project-memory and conversation-summary controls, native project-memory build hooks, CodeGraph/pstack dependencies, and OpenCode intelligence operations.
• Retain the upstream deletion of the obsolete client AVC repair code.
• Enable usable Windows testing in the fork: retain the upstream Blacksmith runner for pingdotgg/t3code, use GitHub-hosted windows-2025 for forks, and install each owning package's lockfile-pinned Chromium shell for web/server tests. Build the server's native addon before file-scoped tests, matching the prerequisite already present in the full server test command. Run each non-server package independently, attempt all 34 suites, and fail the overall job if any suite fails. The trigger remains manual, matching upstream; the 45-minute job limit and test assertions/timeouts are unchanged.

Validation

Local validation and all non-server Windows suites are complete, with failures recorded below. Windows is advisory: 403 of 474 server files have audited results, and the final 71-file diagnostic group continues without blocking merge readiness. Current PR head is 7b90da3e81073a49db88089f1d22e84ab68d754a, tree a25d8801e459f43cfbeda0cf05cfe9392e450cfe.

Full local application suites ran on e0149fb. The three subsequent commits change only .github/workflows/windows-tests.yml; application, test, and dependency inputs are unchanged. All three workflow changes independently passed vp run check:changed HEAD --concurrency 1 (40 typechecks, formatting, project-memory addon build, and Agent Context Budget). The newest gate passed in 19.03 seconds outside the restricted sandbox after its initial restricted invocation stopped before test output; source, command, worker, heap, and workflow hash were unchanged. Each published tree was compared with the exact locally validated tree.

Merge policy

Windows Tests is optional diagnostic coverage, as requested by the user. Queued, running, failed, or cancelled Windows runs do not block merge readiness. The workflow stays enabled and manual, and its actual failures and coverage limits remain recorded below. The fork's Fork Check aggregate already excludes Windows, and the repository has no classic branch-protection rule or ruleset requiring this workflow. The signed-in PR page also exposes an enabled merge action while the Windows run is active. Existing T3 merge monitors now explicitly exclude only .github/workflows/windows-tests.yml from their waiting/green criteria; their non-Windows validation, current-head verification, review, and protection requirements are unchanged. This exception does not apply to Windows-related assertions within the Linux Client and package tests job. No workflow code change or CI restart was needed for this policy change.

Completed results

Check	Result
All 34 non-server workspace package suites (local Linux)	16,146 passed, no failures, across 1,426 passed files. Three original declared skips; the optional Nushell case subsequently passed, leaving two platform/UID skips.
Full web suite, included above	5,483 passed: 4,766 unit tests and 717 Chromium browser tests.
Full mobile JavaScript/React suite, included above	3,238 passed: 2,959 unit tests and 279 React tests.
Full local server suite	All 6,511 tests / 474 files accounted for. Initial run: 6,493 passed, 11 failed, 7 skipped. After focused environment corrections: 6,498 passed, 9 failed, 4 intentional skips across unique cases.
Extra web render scenarios	11 passed / 1 failed. The same failure reproduces on exact fork main. The three-pass wrapper stopped after its first failed pass, so repeatability is not established.
Production web boot	Passed: the built production chunk graph loaded and mounted React in Chromium.
Repository formatting and typechecks	Passed: 6,740 formatted files and all 40 workspace typechecks.
Full lint	11 existing errors in seven unchanged files; exact baseline reproduces all 11.
Knip and generated GraphQL	Passed: both knip passes and both generated GraphQL checks.
Production builds and smoke checks	Passed: web/server/desktop pipeline, preload verification, release smoke, marketing build, standalone CLI executable, and Node/executable help/version smoke checks.
Rust and native binding	49 Rust tests passed across four crates; all four Rust format checks passed; release binding build and 6 Node binding tests passed.
Workflow helpers	53 passed: 47 Node and 6 Python cases.
Mobile native static analysis	Passed: SwiftLint, ktlint, and detekt.
Android standalone and generated-project checks	17 standalone review-diff cases passed; 138 generated-project JVM/Robolectric test executions passed across five module suites and configured Android API versions; notifications lintRelease passed with 0 errors and 15 warnings. The repository-documented UAST flag was supplied; current lint reports that the flag no longer has an effect. Six review-diff cases were exercised by both harnesses and are not additional unique coverage.
Additional regression checks	Python seed regression 4/4 passed; optional Nushell login-shell probe passed.
Advisory mobile fingerprints	Identical on PR and exact fork baseline for iOS (118 sources) and Android (120 sources), with no source/hash differences. This check identifies no new native-build requirement from the PR.

The reconciled server figures combine the initial complete run with focused follow-ups; they are not a second full-suite run. Baseline reproductions, CI repetitions, and resource-interrupted attempts are not counted as additional unique passing tests. The large local Vitest suites used one worker to fit the shared host; no test-source changes, timeout increases, or new skips were introduced.

Remaining failures and baseline evidence

• Lint: t3code/no-unreasoned-state-sync reports 11 errors in ProjectKnowledgeFactEditor, ProviderAuthenticationSection, CodeExplorerDialogHost, LoomTimeline, useLoomViewGestures, LoomMinimap, and LoomWeaveField. The seven files, lint configuration, and custom rule implementation are unchanged; the same errors reproduce on exact fork main e747cef.
• Foundation persistence: starts later thread work behind a delegated-task stop that never returns (replayed: false) fails identically on PR and exact baseline. The complete file reports 46 passed / 1 failed on both. The precise virtual-clock scheduling mechanism remains unproven.
• Muse replay: eight PR scenarios fail. The representative fast simple/muse case reproduces identically on baseline; expanded logs show the same additional 571-byte <github_reads> block. All eight recordings omit that block, and relevant implementation, fixtures, and dependency inputs are unchanged. All eight baseline scenarios were not individually rerun.
• Render count: app-stream-assistant-turn at apps/web/test/appScenarios.renders.tsx:210 expects 30 MarkdownParagraph renders and observes 59. The exact baseline produces the identical result.
• The original Linux CI server shard 4 independently reported the same Foundation failure and eight Muse failures: 762 passed / 9 failed.
• Two initial theme permission failures were artifacts of UID 0 bypassing chmod; the complete seven-test file passed with the two filesystem bypass capabilities removed. Three opt-in HtmlRender Chromium cases also passed. Those browser cases verify application restrictions, not Chromium's OS sandbox.

Windows and latest CI

Windows Tests is enabled, as it is upstream. The first two runs were cancelled because the fork had no Blacksmith integration or matching runner. The first hosted non-server attempt completed only pstack (41 passed / 3 failed) before Vite+ cancelled the other started packages. Its executable-flag failures are in a pstack tree byte-identical to exact fork main; no separate Windows baseline execution was performed. The checkout exit-128 cleanup message was a warning, with the cleanup step marked successful.

The follow-up workflow change attempts each of the same 34 packages and retains a failing final exit when any package fails. Discovery was checked against the complete local inventory with no missing, extra, or duplicate paths.

• Full non-server Windows run on 2672d23: completed with failures, without reaching the job timeout. All 34 packages were attempted and emitted final summaries: 19 passing / 15 failing packages, 15,940 passed / 102 failed / 98 skipped tests, and 1,384 passed / 34 failed / 9 skipped files. The Test step took 23m 34s.
• Initial Windows server run on 685d77a: cancelled at the 45-minute job limit, confirmed by GitHub's execution-time annotation. It emitted terminal results for 190 of 474 files: 167 passed / 22 failed / 1 wholly skipped files and 4,758 passed / 142 failed / 91 skipped tests. These are exact counts for the result-bearing files and lower bounds for the interrupted run; there was no final Vitest aggregate. 284 files had no completion result; the log does not prove whether each had started. All 22 failing test files are unchanged from fork main, which is input-equivalence evidence rather than a separate Windows baseline execution.
• Four bounded server follow-ups on 7b90da3 target precisely those 284 files plus ServerBrowserPage.test.ts, whose original suite setup left 18 cases skipped. The new workflow's Chromium installation and native-addon build have passed in all four jobs. The manifests contain 285 unique files with no omissions or duplicates; only the browser file overlaps the original 190 result-bearing files. Group 1 completed all 72 files: 68 passed / 4 failed files; 422 passed / 7 failed / 3 skipped tests. Group 3 completed all 71 files: 66 passed / 5 failed files; 343 passed / 10 failed / 1 skipped tests. Both emitted selectors exactly match the manifests and their per-file counts reconcile with final Vitest summaries, with no unhandled-error section. Group 2 also completed all 71 files, with 367 passed / 11 failed / 3 skipped tests and an exact selector/aggregate match. ServerBrowserPage passed 22/22, replacing its original 4 passes and 18 setup skips. The 11 failed cases comprise eight Muse replay cases, two Grok text-generation timeouts, and one DatabaseMaintenance EBUSY cleanup failure; a separate Muse cleanup-hook error is not counted as an additional test case. These runs now account for 403 of 474 unique server files. The independently audited interim union is 403 files: 369 passed / 33 failed / 1 wholly skipped, with 5,886 passed / 170 failed / 80 skipped cases (6,136 outcomes). The original browser row is replaced exactly once. These are reconciled observations from multiple runs, not complete full-suite totals. Only group 4: 71 files remains active and uncounted in this audit. Its Chromium/native prerequisites passed; the last captured Test status was in progress at 01:44:40 UTC on November 1. It continues as an optional diagnostic, with the existing 45-minute job limit, and does not hold merge readiness. No additional Windows reruns or weaker test assertions are part of this policy.
• Current-head Agent Context Budget passed on 7b90da3; it reported that agent context is not larger than main, with no increase-label exception. Fork CI has passed change detection; 11 downstream jobs remain pending at this update (11 queued, 0 in progress). Mobile native checks was skipped by change detection. Superseded PR runs were automatically cancelled by the existing concurrency policy. Earlier checks/build and Rust successes remain historical results and do not substitute for the current head's CI status. Expanded local native validation is recorded above.

The non-server Windows failures include CRLF/path expectations, executable-bit assumptions, socket/shell fixtures, SQLite cleanup, browser assertions/timeouts, and the payload budget. 33 of the 34 failing test files are byte-identical to exact fork main; the one changed file is the packaging test discussed below. This is input-equivalence evidence, not proof that every failure reproduces on a Windows baseline: application inputs such as ChatView changed. Three scripts fixtures failed because git init could not access \\.\nul; these are actual Test failures and are separate from the warning-only Post Checkout submodule cleanup. The Windows-only npm-shim case passed.

The conflict-resolved model-picker interaction tests passed on Windows (2/2, plus 2/2 composition cases). The conflict-resolved cross-architecture executable-probe case also passed. A separate packaging fixture exceeded its existing file limit: 2,882 files versus 1,800; the failing test block, production packaging implementation, pstack sources, and dependency inputs are unchanged. The precise contents responsible for the excess are not inventoried in the log, and no separate Windows baseline was run.

The isolated Windows ChatView follow-up passed 1/1 on the same 2672d23 commit, with the case completing in 8.406 seconds under the unchanged 15-second timeout. The full-run timeout did not reproduce in isolation; its runtime cause remains unproven. Source review found no direct connection to the removed thread-family subscription, which was already inactive for this draft route. The original full-run figures above are retained separately; this is not a second full-suite run.

Platform and environment limits

• icons:check requires Apple Icon Composer/Xcode on macOS and could not run on Linux.
• The iOS review-diff runtime harness requires Apple SDKs and UIKit/Mac Catalyst. SwiftLint success does not establish iOS runtime coverage.
• Declared mobile device journeys require a configured Simlock pool and the required Tailscale/device environment, which were unavailable. No mobile device journeys were run.
• The repository's integrated test-t3-app / test-t3-mobile skills require T3 Browser/Device tools absent from this session; live-client visual captures were unavailable. Automated Chromium coverage and production boot passed.
• Remaining local declared skips are platform/UID/version conditions: a Windows npm-shim case, a Windows uninstaller integration, three UID-0 permission cases, and an Undici-7-only workaround (the pinned runtime uses Undici 8).

Original merge validation

Before opening the PR, VITEST_MAX_WORKERS=1 NODE_OPTIONS=--max-old-space-size=2048 vp run check:changed origin/main --concurrency 1 passed: 215 tests across 15 files, 15 React Compiler checks, all 40 typechecks, formatting, changed-file lint, both knip passes, generated GraphQL verification, unused dependency patch checks, and the native addon build. These are subsets of the expanded coverage, not additional unique tests.

The account-disclosure regression was reproduced by removing the pooled-account exception, then verified with the corrected interaction test. At that point, Git's merge-tree check also confirmed a clean merge with the then-current PR #36, with no overlapping changed paths.

Model: GPT-6 Astra Pro. Harness: ChatGPT with repository shell, GitHub tools, and independent review agents.

[changeset-bot]

⚠️ No Changeset found

Latest commit: 7b90da3

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR