t3x sync 2026-08-10: 43 upstream commits onto the fork's 91 patches

[radroid]

Rebases the fork's 91 patches onto upstream/main = 3d74474f6, absorbing 43 upstream commits (147 files, +7272/-1497) from merge-base a20923ce4.

Important

Do not use the merge button. This branch is the fork's series rebased onto new upstream, so origin/main is not its ancestor — GitHub shows it as conflicting and merge/squash/rebase all fail. Land it with:

git fetch origin && git push --force-with-lease origin origin/t3x/sync-20260810:main

Recovery point: tag t3x/pre-sync-20260810 → 7bcf71ee3 (pushed to origin).

The headline: zero conflicts, and the ledger did not move

All 91 fork patches replayed with no conflicts, and the seam ledger's file-list half came out byte-identical — same 37 files, same +1957/-912.

That is the result, not a skipped step. A clean rebase is exactly when this fork's silent-drift hazards bite, because nothing forces anyone to look at the merge. So each was checked explicitly.

What was audited, and why

Seam deltas (the regression detector). Upstream touched five seam rows this range — ChatView.tsx, SettingsPanels.tsx, ChatComposer.tsx, contracts/src/settings.ts, AppSymbol.tsx — and every one's fork delta came out unchanged. That means the fork's additions landed on top of upstream's new content rather than displacing it. A resolution that silently re-took fork content over an upstream change shows up here as a row whose delta grew while nobody edited that file.

Logic mirrors. Upstream moved four mirror sources: decider.ts (3 commits), http.ts, ProjectionSnapshotQuery.ts, ClaudeAdapter.ts (2). None of the specific mirrored functions changed — hasOpenBlockingRequest, authenticateRawRouteWithScope, getSnapshot/getCommandReadModel, and the mid-turn steering behaviour are all untouched. The mobile-outbox mirror is unaffected too; the only change under apps/mobile/src/state/ is a new usage.ts.

The one that looked dangerous. e70cdb478 "stop Claude resume handshakes from completing turns that never ran" (pingdotgg#5710) removes an untargeted turn.completed, which sits close to the fork's auto-resume. It does not reach it: the fork's reactor triggers on account.rate-limits.updated with status: "rejected", not on turn lifecycle. The change makes thread state at decision time strictly more accurate.

Parallel paths. No send-gating change anywhere in the range — sendDisabledReason, hasSendableContent, canSubmit are all untouched, so the fork's queue path is not bypassing a new upstream guard. Upstream's new unsent-drafts feature (pingdotgg#5777) is not a second send path: composerDraftStore.ts is pristine upstream with zero fork delta, and pingdotgg#5777 only surfaces existing drafts in the sidebar.

Fork hoists. ComposerPrimaryActions.tsx — which holds the fork's hoisted send/stop markup, and which silently reverted an upstream restyle two syncs ago — had zero upstream commits this range.

The fork's settings row. Upstream overhauled project settings (pingdotgg#5768, pingdotgg#5923) and added breadcrumbs (pingdotgg#5929). The fork's "Notify when an agent needs input" row is in GeneralSettingsPanel, still mounted by routes/settings.general.tsx; the overhaul was a different route.

Dependencies: the 08-08 failure mode cannot arise here

Upstream changed no dependency. The only manifest edits in the range are exports-map entries in packages/shared (threadEnvMode, usageMerge, usageFormat) and one script alias in the root package.json. So pnpm-lock.yaml needs no regeneration, and the security sweep carries over untouched rather than being re-derived. Dependabot should stay at 6.

Verification

• vp run typecheck — exit 0, no errors
• vp run lint — exit 0 (only the pre-existing upstream ThemeEditorPanel.tsx unused-disable warning)
• vp run test --testTimeout=120000 --hookTimeout=120000 — running
• t3x fork CI on this branch — running

Post-landing follow-up

Two new upstream workflows arrive active and can only be disabled after they register on the default branch:

• .github/workflows/web-preview.yml (ci: label-gated hosted-web preview deploys pingdotgg/t3code#5465)
• .github/workflows/mobile-fingerprint-check.yml

Both are pull_request-only, so neither can fire on a force-push landing. mobile-eas-production.yml was modified by pingdotgg#5609 but stays disabled_manually.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 945f3cdc-642c-4429-badb-b139f856a3c5

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[radroid]

Local full suite: green, exit 0. All 15 packages, no failures, no retries.

2 · 4 · 17 · 5 · 41 · 4 · 1 · 4 · 17 · 46 · 68 · 101 · 27 · 234 · 227 (+2 skipped)

Worth flagging because it is unusual: the previous two syncs each failed the batch run on a single, different upstream-owned file (nativeReviewDiffHighlighter.test.ts, then MessagesTimeline.test.tsx), both passing standalone. Those were machine contention — 15 packages in parallel on a 16GB Mac — not regressions. This run had the machine to itself and came back clean first time, so there is no flake to explain away and no per-package re-run needed to trust it.

(One of those is worth remembering regardless: MessagesTimeline.test.tsx has beforeAll(async () => {…}, 30_000) as an in-file literal, which overrides --hookTimeout entirely. No CLI flag can raise it.)

Waiting on t3x fork CI, which is the authoritative gate.

[radroid]

t3x fork CI: green (run 31354783139) — verify succeeded on e4fc808df, the branch tip.

Everything is now green: typecheck, lint, local full suite, and fork CI. Ready to land whenever you've reviewed.

Landing is a force-update, not the merge button:

git fetch origin && git push --force-with-lease origin origin/t3x/sync-20260810:main

Recovery point if anything goes wrong: git push --force origin t3x/pre-sync-20260810:main (tag → 7bcf71ee3).

Two post-landing steps remain, both mine to do once it lands: disable web-preview.yml and mobile-fingerprint-check.yml (they can only be disabled after GitHub registers them on the default branch), then confirm Dependabot is still at 6 and watch the release.

[radroid]

Rebuilt — the conflicts were real, and not the artifact I described earlier

I was wrong to attribute the CONFLICTING badge to the usual rebase-onto-upstream artifact. It was a genuine collision, from a cause I hadn't considered:

The scheduled t3x upstream sync (daily) workflow force-landed its own rebase onto main at 09:24, while this PR was open. It rebased the same 91 fork patches onto a newer upstream base (9821bca1c). So main and this branch held near-identical content under completely different SHAs with no shared history — which GitHub can only report as unresolvable. No amount of conflict-resolution on the old branch would have fixed it; the branch had to be rebuilt from the new main.

What changed

Branch force-updated e4fc808df → 4882a70c2:

• Rebuilt from the post-automation main and rebased onto upstream/main = 78f462c4e (upstream v0.0.33), absorbing 5 further commits the automation missed — the release prep, three mobile fixes (fix(mobile): parse EAS fingerprint JSON pingdotgg/t3code#5991, fix(mobile): stop Android user bubbles with code blocks from overlapping pingdotgg/t3code#5659, fix(mobile): detect PowerShell cmdlet errors in work log rows pingdotgg/t3code#5726), and fix(web): use import/export-appropriate icons for theme buttons pingdotgg/t3code#5964.
• Ledger re-baselined against 78f462c4e. Worth noting: after the automation's landing the ledger header still claimed merge-base a20923ce4 — two baselines stale. The automation lands without touching the ledger, which the ledger's own self-reference rule forbids.

Audit, re-run against the new base

• Seam surface unchanged again: 37 files, +1957/-912 — byte-identical, now across two consecutive rebases and 49 upstream commits.
• The check that made this cheap: every fork edit to a shared file is additive (preload.ts +29/-0, channels.ts +9/-0, ipc.ts +44/-0, __root.tsx +9/-0). Anything displacing upstream content moves a deletion count. Concretely — fix(web): use themed confirmation dialogs pingdotgg/t3code#5624 removed a line from channels.ts and preload.ts, and that removal survived, because the fork appends rather than re-stating.
• apps/desktop/package.json is the sole +1/-1 row and carries both sides: upstream's 0.0.33 bump and the fork's electron 41.10.3 pin (upstream is on 41.5.0).
• Dependabot: still 6. No dependency moved in the range.

Already done

Both new upstream workflows arrived active on main, as predicted, and are now disabled_manually:

• .github/workflows/web-preview.yml
• .github/workflows/mobile-fingerprint-check.yml

Neither had fired.

typecheck 0 · lint 0 · CI + full suite running on 4882a70c2.

[radroid]

Local full suite on 4882a70c2: green, exit 0. 15/15 packages, no failures.

File counts moved as expected for the 5 newly absorbed commits — apps/mobile 101 → 102 and apps/web 234 → 235 test files — which is upstream's own new coverage arriving, not fork churn.

2 · 4 · 17 · 5 · 41 · 4 · 1 · 4 · 17 · 46 · 68 · 102 · 27 · 235 · 227 (+2 skipped)

Waiting on fork CI, then landing.