chore(upstream): adopt the 2026-08-18 T3 Code batch

[rynfar]

Adopts all eleven upstream commits from bab4b6f02..82b8a9380 (2026-08-17 – 2026-08-18). None were patch-equivalent to Pylon, and the deferred register was empty going in and stays empty.

What's here

#6466 protects source-control API budgets — the most valuable change in the batch, and it applied without a single conflict. One GitHub pull request detail load was spending 104 GraphQL points: it asked for 100 replies per review thread, followed every reply cursor, and refreshed every minute. This adds an Effect cooldown service keyed by provider plus host, native rate-limit detection for GitHub, GitLab, Bitbucket, and Azure DevOps, a GraphQL cost budget reserving the last 10% for interactive actions, and 10-reply pagination behind Load more comments. Upstream measured 104 → 14 points on a 44-thread PR. Three behavior changes ride along and were accepted deliberately: long threads paginate instead of loading eagerly, live refresh slows 1 min → 5 min, and the idle cutoff moves 5 min → 6 min.

#7172 stops threads lying about being busy — a late task.progress with no status was re-adding an already-idle task to the live set, leaving the sidebar on Working with zero live agents.

#7082 + #7083 add Settings → Integrations — default preview viewport, zoom, appearance, and floating-preview auto-show, plus an Agent browser access toggle that withholds the MCP credential in prepareMcpSession. That's the only place a credential is minted, so one branch covers all five providers.

#7209 bans native title tooltips behind a new error-level oxlint rule and migrates 33 sites. #7077 makes review verdicts a first-class timeline row. #7219 removes a silent .slice(0, 8) that was hiding up to 82 rows of the usage breakdown. Plus #7321, #7276, #7296, and #7364.

Pylon-first resolutions

Applying in upstream chronological order mattered: #7083's conflicts and #7077's both disappear once their predecessors land first.

• ServerUpdateAction.tsx — kept Pylon's <DotMatrix state="error"> where upstream has a plain destructive dot; took the tooltip wrapper.
• Sidebar.tsx — kept Pylon's themed text-warning token over upstream's text-amber-700 dark:text-amber-300 on the Woke pill.
• BrowserDeviceToolbar.tsx — upstream's cn() guards are constant inside each ternary branch, so Pylon's plainer classes say the same thing; kept ours.
• CodexSessionRuntime.test.ts — took upstream's new codexDefaultModeDeveloperInstructions(hasBrowserTools) function, kept Pylon's /Pylon/ assertion.
• SettingsPanels.logic.test.ts — both sides append describe blocks at the tail; both kept.

Two fixes beyond conflict resolution: #7083 asserted the absence of "T3 Code collaborative browser", a heading Pylon renamed long ago, so it passed vacuously — repointed at Pylon's actual heading. And the new Integrations panel's module comment named T3 Code.

Because #7209's rule lands at error, upstream's migration isn't sufficient on its own — running the rule found four more sites in ProviderUsageMatrix.tsx, a Pylon-only surface. Three became styled tooltips; the row-label cell never truncates, so its title only repeated visible text and was dropped.

Verification

Typecheck clean across contracts, shared, client-runtime, web, desktop, server, and mobile. Targeted tests: server 35 files / 707 tests, web 32 / 245, desktop 1 / 7, oxlint plugin 1 / 11 — all passing. vp lint reports 0 errors over the touched trees; vp fmt --check clean. routeTree.gen.ts is generated, so its imports were checked against the route files on disk rather than trusted.

Not verified in a real client. The surfaces that would most repay a pass are Settings → Integrations, the pull request detail, and the two mobile changes.

Built with Claude Opus 5 (1M context) in Claude Code.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for 68823f1.

This comment will update automatically after the next completed run.

[rynfar]

Review pass

An xhigh review plus a browser pass against the branch. Both confirmed the integration; neither is a blocker.

Verified live (seeded database, real GitHub data): the new Settings → Integrations page lists in the nav, with #7083's Agent browser access enabled and #7082's four client-local defaults dimmed as one block — the intended split. #7209 checked at runtime rather than by lint alone: zero native title attributes remain on intrinsic elements in the rendered DOM. #7077's Reviewers row renders above the conversation.

Not reachable: #7219's usage breakdown never finished loading against the seeded database, and no pull request in reach carried an approval or an 11-comment thread. Mobile was not exercised.

All five Pylon-first resolutions confirmed clean, including the BrowserDeviceToolbar equivalence claim — upstream's cn(aspectRatio !== null && …) guards really are constant inside each ternary branch.

Two inherited upstream defects found, deferred to a follow-up as DEF-3 and DEF-4 so this PR stays a faithful adoption. Both exist in upstream main:

1. #6466 narrowed truncated to "a thread has an 11th reply" while three readers still treat it as "the conversation is short of the host" — #7077's approval badge is suppressed, the handoff prompt gains a spurious notice, and the agent prompt receives only a thread's oldest 10 comments.
2. githubGraphQlBudget compares GitHub's true remaining against a locally-reserved figure, so honest responses are always discarded and the cost estimate only ratchets upward. Fails safe and resets hourly.

One finding dismissed: the #7172 liveness guard matches the module's documented "orphaned background work is not live" intent.

Details in .agents/upstream-review.md (68823f1).