Skip to content

chore(upstream-sync): port 29 t3code commits - #452

Merged
trancong12102 merged 32 commits into
mainfrom
chore/upstream-sync-20261120-0233
Nov 19, 2026
Merged

trancong12102 merged 32 commits into
mainfrom
chore/upstream-sync-20261120-0233

Conversation

@trancong12102

@trancong12102 trancong12102 commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

Ports 29 upstream t3code commits in two bundles: 1852121afa..42d2387285 (6 taken, 2 skipped) and 910c1a8df9..0f32645852 (20 taken, 1 skipped). The first bundle:

  • background git status fetches no longer leave partial packs
  • ACP sessions keep their model controls when a session/load reply arrives late
  • Ctrl+B reaches a focused terminal
  • right-click opens thread actions on sidebar search results
  • a client-runtime perf shortcut
  • docs for thread launch retries

The second bundle:

  • long user messages fold on mobile like on web
  • Find in Thread matches dotted capital I and scrolls wide tables to the match
  • whole-word workspace search no longer misses files
  • HTML reports in dot-prefixed folders open
  • native session replacement fails honestly when a provider reuses its session
  • repeated checkpoint capture timeouts stop delaying turn starts
  • restart recovery stops reading every idle thread on a shared session
  • remote task results skip discarded history
  • the dev server stops restarting when something writes under packages/*/node_modules
  • smaller fixes: account badge initials, delegated wakes name the child thread, preview_resize on hidden desktop tabs, ACP image previews, Antigravity compaction, the SSH missing-tar message, no Git spawn in a removed worktree, a quieter projection read span, triage filing, Simplified Chinese strings

Risky commits

  • 1852121afa fix(server): failed background status fetches no longer leave partial packs (shared core VCS, every forge).
    • The status fetch now downloads with a dry run into a private object directory, publishes those objects, and then moves refs with a sealed second fetch.
    • Upstream added its own adoptPrivateObjects to publish the objects. It duplicated Onicode's checkpoint object publication with weaker guarantees:
      • new fanout directories ignored core.sharedRepository
      • a .pack could land before its .idx
      • an incomplete pack pair was moved
    • Adapted (8b313bc): Onicode's publisher moved into apps/server/src/vcs/gitObjectPublication.ts as publishPrivateObjects. Checkpoints and the status fetch both use it, and upstream's copy is gone.
    • The port script renamed the quarantine prefix to tmp_objdir-onicode-.
  • 5f8ab374af fix(acp): a late session/load response no longer leaves the session without model controls (provider, 837 lines). ACP-based providers only. Taken as is; the T3 mock env names were renamed to Onicode's. ACP adapter and JSON-RPC suites pass.
  • 42d2387285 fix(web): right-clicking a sidebar search result now opens its thread actions (sidebar). Taken as is. It only adds one useStableCallback after handleThreadContextMenu, so Onicode's hook order is untouched. "does not re-render thread rows when the sidebar renders with nothing changed" passes, and so does Sidebar.drag.browser.test.tsx.

Second bundle:

  • 910c1a8df9 fix(mobile): long user messages collapse like on web (mobile). Taken as is. It is the mobile half of the existing web fold, and its strings are in all 10 locales. The fold measures once per layout and draws its gradient only while folded. Mobile typecheck passes, and so does filePreview.test.ts.
  • 7922611cee fix(i18n): Simplified Chinese strings (mobile). The marketing zh.ts hunk is dropped, because Onicode has no marketing site. The app strings are taken.
  • 44e3f42e69 fix(server): a native subagent start no longer records a failed projection read span. CONFLICT in ProviderEventIngestor.ts, resolved by hand. The lookup uses the new findThread, and Onicode's /rename title-ownership logic around it is kept (upstream has nothing like it).
  • d32df2d34c fix(server): native session replacement fails honestly when the provider reuses its session (shared core turn start, 410 lines).
    • There was a CONFLICT in the ProviderSwitch.integration.test.ts fixture options; both Onicode's holdRollback and upstream's beforeForcedNativeThread are kept.
    • Onicode's own test "keeps a scheduled fresh session's thread apart … on a provider restart" timed out on the new reused-session check. Its stand-in Codex adapter returned the same native id when asked for a fresh session. It now gets a new id per replacement, as real Codex does.
    • Claude and Codex clear nativeThreadRef on replacement, so the new error fires only for providers that really reuse their session.
  • 1e70210250 perf(server): remote task results skip discarded history (linked environments, tagged auth). Taken as is. It relies only on thread.itemCount, which the contract already requires, so older linked environments still work.
  • 1e9f4c45c3 fix(triage): file labeled reports through the issue template (ci). Taken. Onicode's "Triage report" template exists. Note: the repository has no via-triage label yet, so the template cannot apply it (this predates the sync).
  • 0f32645852 fix(dev): avoid dependency-triggered server restarts (deps). See the adopt-or-keep decisions below. CONFLICT in apps/server/package.json: Onicode has no dev:built, so that line is left out.

Routine commits

  • 21df2269bd docs(mcp): document thread launch retries. Taken as is (oni_thread_launch takes clientRequestId in Onicode).
  • c8625cafce fix(web): Ctrl+B reaches a focused terminal again. Taken as is.
  • ff4197c480 perf(client-runtime): skip history scans when no runs are queued. Taken as is.

Second bundle, taken as is unless noted:

  • 3c08684f1d fix(client-runtime): CamelCase account names get distinct badge initials
  • 5533151d77 fix(server): delegated completion wakes name the child thread
  • 4e39664248 fix(server): preview_resize sizes hidden desktop tabs without a snapshot
  • 0240eac9da fix(server): images that ACP agents read show a preview
  • cdf5a6c0fc fix(antigravity): allow compaction when the runtime advertises it
  • 031dd1a818 fix(ssh): report missing tar before downloading the remote runtime
  • a4d0151700 fix: match dotted capital I in Find in Thread (scenario adapted)
  • 55e379bb96 fix(server): the HEAD watch no longer spawns Git in a removed worktree
  • 9f4d25495d fix(server): whole-word content searches no longer miss matching files
  • 5bc2d512e4 fix(web): scroll wide tables to the selected Find match (scenario adapted)
  • 61085457d7 fix(server): repeated capture timeouts no longer delay turn starts
  • 800b0748bc fix(server): restart recovery no longer reads every idle thread on a shared session
  • 6d05c5f496 fix(server): HTML reports in dot-prefixed folders open (scenario adapted)

Adopt-or-keep decisions

  • Kept Onicode's checkpoint object publication over upstream's new adoptPrivateObjects. Onicode's version preserves shared-repository fanout permissions, links metadata before the pack, publishes only complete pack pairs, and treats an existing object as success. The status fetch now reuses it.

  • Kept Onicode's DevToolsActivePort wait over upstream's reserved port (2ab3b7f59c). Onicode's 578ef81 waits for the port Chromium actually bound. Upstream releases a reserved port before launch, so another process can take it in between.

  • Adopted upstream's dev watcher over Onicode's node --watch-path=../../packages (0f32645852).

    • I first skipped it as Windows-only. Claude's review pointed out that --watch-path=../../packages also covers packages/*/node_modules.
    • A probe confirmed it on macOS: Node's watch restarted on a write to node_modules/.vite/results.json. Vitest's cache and vp i write there, so they restarted vp run dev:server.
    • Upstream's apps/server/scripts/dev.ts watches only each package's src and package.json. A probe with the new script ignored the node_modules write and restarted once on a src change.
    • Codex suggested keeping Node's watch on macOS and Linux, but the probe shows the same bug there.
  • Kept Onicode's /rename title handling in ProviderEventIngestor (44e3f42e69), since upstream has no equivalent, and took upstream's findThread lookup.

Skipped

  • d3da5c6f32: marketing homepage copy only.
  • 2ab3b7f59c: test race already fixed in Onicode (see above).
  • cb2a488c67: off-track (nightly/stable) desktop build warning. Onicode ships one channel through better-update and has no desktopUpdateChannelOf or channel switch.

E2E scenarios

Three ported to the desktop target, each pressing ControlOrMeta+… like the other desktop scenarios. All three pass locally with vp run scenarios --target desktop:

  • scripts/scenarios/thread-find-matches-dotted-capital-i.e2e.ts (a4d0151700)
  • scripts/scenarios/find-reveals-clipped-table-cell.e2e.ts (5bc2d512e4)
  • scripts/scenarios/html-report-in-dot-folder.e2e.ts (6d05c5f496)

None were deferred.

Follow-ups

Queued in the follow-up queue (## From sync chore/upstream-sync-20261120-0233). The batch fix thread Batched fixes: Force stop/session release, PR watch, mobile RTL… (queue 2026-11-20 00:35) is still running, so no new thread was started.

All of these are upstream code as written:

  • P1 (1852121afa): over a dumb (static) HTTP remote that moves between the two fetches, the sealed fetch can store the new tip as a loose object and move the ref. Cleanup then deletes that object.
  • P2 (1852121afa): the expected sealed-fetch race is recorded as a failure and backs off for up to 15 min.
  • P2 perf (1852121afa): every uncached status refresh runs git count-objects -v only to read the alternates.
  • P3 (c8625cafce): explicit Ctrl bindings for sidebar.toggle with when: terminalFocus never fire.

From the second bundle, queued under ## From sync chore/upstream-sync-20261120-0233, second bundle. The same batch thread is still running, so no new thread was started. All of these are upstream code as written, and Codex raised all of them:

  • P1 (6d05c5f496): an already-issued HTML preview URL re-resolves its folder on each request. If .reports is later swapped for a symlink to another hidden folder, the URL serves that folder's previewable files.
  • P2 (9f4d25495d): 100 lines with éneedle fill the per-file cap, so a later standalone needle in that file is skipped.
  • P2 (4e39664248): a paint capture still pending when the resize wait ends keeps a hidden desktop tab painting and queues later screenshots behind it.
  • P2 (5533151d77): remote follow-up wakes still carry no thread identity.
  • P2 (910c1a8df9): a mobile search jump lands on a long user message but leaves it folded.
  • P3 (910c1a8df9): a folded mobile message keeps a stale height decision after reflow.
  • P3 perf (1e70210250): the remote result tail probe can add a round trip.

Left to upstream (Grok/ACP only), both from 5f8ab374af:

  • A mode notification before a late session/load reply drops the model catalog.
  • A failed candidate load interrupts the active session's pending load response.

Desktop follow-up, not queued: keyboard access (Shift+F10 / ContextMenu key) to a search result's thread actions. Focus stays in the search box.

Review findings

Codex gpt-6.2-astra (high) and a Claude Opus subagent reviewed the same brief independently.

  • Weaker object publisher in 1852121afa. Both raised the shared-repo permissions problem; Codex also raised pack-before-idx.
    • Fixed by reusing Onicode's publisher (8b313bc).
    • The existing checkpoint test "preserves Git's shared object directory permissions" covers the shared function.
  • Dumb-HTTP loose objects in the sealed fetch. Raised by Codex. Queued (P1).
  • Sealed-fetch race backoff. Raised by Claude. Queued.
  • count-objects cost per refresh. Raised by both. Queued.
  • Ctrl terminal bindings. Raised by both. Queued.
  • Grok/ACP late-load edge cases. Raised by both. Left to upstream.
  • Search result keyboard context menu. Raised by Codex. Follow-up above.
  • Both skips. Both reviewers confirmed them.

Second bundle (head db661c6, then 58b818b for the watcher), reviewed independently by Codex gpt-6.2-astra (high) and a Claude Opus subagent on the same brief:

  • Neither found a bug introduced by the ports or their adaptations. Both checked the /rename conflict resolution, the fixture change and the scenario key presses.
  • 0f32645852 skip. Claude said the reason was incomplete and Codex suggested a Windows-only take. Verified with probes and taken in full (above).
  • cb2a488c67 skip. Both confirmed it.
  • Upstream-as-written findings (Codex): queued, listed under Follow-ups.
  • Checkpoint 5 s baseline wait on slow repos (Claude): not a regression. No action.
  • Ledger rows repeat commit titles (Claude, nit): dismissed. That is the port script's default note.

Checks

Second bundle: vp run check:changed 1303c1d45c passes (typecheck of 41 packages, 27 test files, React Compiler check), and so do vp run i18n:check, the wide-table Find cases in MessagesTimeline.browser.test.tsx, the ProviderSwitch, ThreadStop and ThreadManagementService suites, and the three scenarios.

First bundle: vp run check:changed passes except three cases in apps/server/src/vcs/GitVcsDriver.test.ts: "checkpoint bulk staging cleans operation-owned packs after interrupt/timeout/success and can retry". They fail the same way on this macOS host with origin/main's VCS sources (ENOENT reading the private pack directory), so they are unrelated to this sync. CI runs them on Linux.

Done with Claude Opus 5.5 in Claude Code, reviewed by Codex gpt-6.2-astra and Claude Opus 5.5.

🤖 Generated with Claude Code

juliusmarminge and others added 30 commits November 20, 2026 02:36
Ported from pingdotgg/t3code@21df2269bd.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
Onicode already fixed this race by waiting for Chromium's DevToolsActivePort file.
Ported from pingdotgg/t3code@c8625cafce.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
…519)

Ported from pingdotgg/t3code@ff4197c480.

Co-authored-by: SAPHID <saphid@users.noreply.github.com>
Co-authored-by: maria-rcks <maria@kuuro.net>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
… actions (#31545)

Ported from pingdotgg/t3code@42d2387285.

Co-authored-by: maria-rcks <maria@kuuro.net>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
…ts do

Follow-up to the port of pingdotgg/t3code@1852121afa: upstream's new adoptPrivateObjects duplicated Onicode's checkpoint object publication with weaker guarantees (fanout directories ignored core.sharedRepository, a pack could land before its index, an incomplete pack pair was moved). Both now share publishPrivateObjects.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ported from pingdotgg/t3code@910c1a8df9.

Co-authored-by: maria-rcks <maria@kuuro.net>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
…als (#35343)

Ported from pingdotgg/t3code@3c08684f1d.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
…7978)

Ported from pingdotgg/t3code@7922611cee. The marketing site change is dropped; Onicode does not ship that file.

Co-authored-by: maria-rcks <maria@kuuro.net>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
Ported from pingdotgg/t3code@0240eac9da.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
…7020)

Ported from pingdotgg/t3code@cdf5a6c0fc.

Co-authored-by: 尤嘉兴 <youjiaxing@dianchu.com>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
…e (#35307)

Ported from pingdotgg/t3code@031dd1a818.

Signed-off-by: bettercallcaleb <121191527+bettercallcaleb@users.noreply.github.com>
Co-authored-by: maria-rcks <maria@kuuro.net>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
…ojection read span (#39284)

Ported from pingdotgg/t3code@44e3f42e69. Onicode keeps its /rename title handling around the new findThread lookup.

Co-authored-by: mInrOz <14320143+mInrOz@users.noreply.github.com>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
Ported from pingdotgg/t3code@a4d0151700. The scenario presses ControlOrMeta+f like the other desktop Find scenarios.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
…e (#39294)

Ported from pingdotgg/t3code@55e379bb96.

Co-authored-by: Martin Bergo <14320143+mInrOz@users.noreply.github.com>
Co-authored-by: Julius Marminge <julius0216@outlook.com>
Ported from pingdotgg/t3code@5bc2d512e4. The scenario presses ControlOrMeta+f like the other desktop Find scenarios.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
…der reuses its session (#27517)

Ported from pingdotgg/t3code@d32df2d34c. Onicode's scheduled fresh session test now gives its alternate Codex adapter a new native id per replacement, as real Codex does; it returned the same id and so tripped the new reused-session check.

Co-authored-by: TinBane <363069+TinBane@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: maria-rcks <maria@kuuro.net>
…29196)

Ported from pingdotgg/t3code@61085457d7.

Co-authored-by: maria-rcks <maria@kuuro.net>
…shared session (#32396)

Ported from pingdotgg/t3code@800b0748bc.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Ported from pingdotgg/t3code@1e9f4c45c3.

Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
…ot Found (#38210)

Ported from pingdotgg/t3code@6d05c5f496. The scenario presses ControlOrMeta+p like the other desktop file-picker scenarios.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: GPT 6.1 Sol <codex@openai.com>
trancong12102 and others added 2 commits November 20, 2026 03:27
Ported from pingdotgg/t3code@0f32645852. Onicode first skipped this as Windows-only, but node --watch-path=../../packages restarts the dev server on macOS and Linux too whenever anything writes under packages/*/node_modules (the vitest cache, vp i). Upstream's watcher watches only each package's src and package.json. Onicode has no dev:built script, so that line is left out.

Co-authored-by: Julius Marminge <julius0216@outlook.com>
@trancong12102 trancong12102 changed the title chore(upstream-sync): port 8 t3code commits chore(upstream-sync): port 29 t3code commits Nov 19, 2026
@trancong12102
trancong12102 merged commit 238d71b into main Nov 19, 2026
27 of 28 checks passed
@trancong12102
trancong12102 deleted the chore/upstream-sync-20261120-0233 branch November 19, 2026 21:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.