[jayleaton]

Brings Ada up to upstream pingdotgg/t3code 580b526758 (last sync was d2c25950b4) and merges current main (#241, #244, #245). Ada's own configuration stays as it was. The upstream merge is a real merge commit (3e39d765fe). Every conflict was resolved by intent, never by blanket ours/theirs.

What stays Ada

• Brand. APP_NAME is "Ada", the iPhone home-screen name is "Ada Agents", and the Pip logo, README and slogan are unchanged. App and bundle IDs are unchanged. 27 new upstream "T3 Code" strings that merged cleanly were caught by the brand guard and changed to APP_NAME/Ada. They were in desktop, server, web and Muse status text, the vite bootstrap title, and the user docs. An orphan "T3 Account" translation key was also removed.
• Repo and releases. CLI_RELEASE_REPOSITORY is jayleaton/ada, and install.sh/install.ps1 are unchanged. The fork nightly and Windows-installer workflows, windows-package-names.ts and the artifact naming are untouched. Upstream's new npx t3@latest hints were replaced, so the SSH tunnel fallback points at Ada's install.sh and the desktop home-owner dialog shows Ada's runnable command.
• Privacy. Telemetry is still off by default, mobile Observe still points at observe.invalid, and the hosted-app default stays opt-in (fix: Ada no longer reports to upstream T3 services #245).
• Product behavior. Notifications are top-level only (fix(web): only top-level chats notify #243). Mobile push goes direct to Expo. Scheduled tasks stay unified. The headless/gateway fixes and Tailscale-only remote are kept. Agent presets are kept, including fix(agents): Captain listeners always run on Haiku #244's Haiku-only Captain listeners. Mobile EAS config and eas.json are unchanged.
• Migrations. Upstream migrations keep upstream ids (through 91). The 24 fork migrations stay on their own ledger. The runner merge keeps the fork's two-ledger schedule and adopts upstream's "started" log line plus one transaction per migration with progress. scripts/upstream-migrations.json is re-synced. A copy of a real fork database (485 MB) migrated cleanly through upstream 82–91.

Conflicts and how they were resolved

• Persistence. The SQLite client takes upstream's writer worker (writerWorker: true replaces the fork's checkpoint worker). The fork's old WAL-checkpoint test is replaced by upstream's read-write worker upgrade test.
• Orchestrator MCP. t3_thread_read takes upstream's beforePosition/earlierPosition in place of the fork's tail. The Claude adapter latch keeps the fork's fix (fix(server): a held wake continuation no longer silences later Claude wakes #98) alongside upstream's Monitor handling.
• Settlement. The fork's subagent settlement rule is kept (upstream fix(server): subagent threads settle with their parent instead of sweeping on their own pingdotgg/t3code#19665 is not adopted). The fork's threadHasPullRequestWatch already covers upstream's PR-watch hold.
• Desktop home-owner dialog. It combines upstream's live owner details (PID, port, version, service vs SSH) with fix: ada command hints follow PATH, escape paths, and fall back to t3 #241's runnable-command hints.
• Tailscale Serve. The fork's port selection and ownership are kept next to upstream's root-holder check and HTTPS probe (fix(desktop): Tailscale HTTPS counts as on only when Serve reaches this backend pingdotgg/t3code#35544, fix(server): t3 pair --tailscale no longer replaces another app's Serve route pingdotgg/t3code#35547, fix(tailscale): Serve setup fails when the tailnet has not enabled HTTPS pingdotgg/t3code#35538).
• Compaction before send. It stays off (fork revert 77245c3 and fix(web,mobile): T3 no longer offers or starts context compaction #87), including on mobile. Upstream's lost-launch-reply outbox tests are kept, and only the compaction block was dropped.
• Thread links. Upstream's structured link regex is used, without the hosted-link alternatives.
• README and release.yml. The README stays the fork's. Upstream's relay public-config job and relay tracing steps are dropped.

Upstream mechanisms adopted because they are better

These are the SQLite writer worker, beforePosition paging, restoreMissingWorktree, listActiveThreads, historical PR handling, PullRequestWatchService, and the live home-owner dialog. Also adopted: useOpenLocalWorkspace, the chunk reload guard, canResumeThread, isServerNetworkAccessible, the channel label helper, the shell's early HTTP load, routeKeepsItsHost, the peer probe, resolveLauncherPath, the tagged-source build version, and the preview listening handshake.

New upstream outbound, telemetry and Connect code removed

None of it is stubbed. scripts/check-removed-connect is green.

• Connect device approval (a2edd28) and everything that depended on it: auth contracts and endpoints, the CLI, the web/mobile approval coordinators and settings, the revoke dialog, clerkLoadFailure, the banners, and migrations 90/91. The migrations stay registered so upstream ids line up, but nothing reads or writes their tables.
• Hosted pieces: Clerk sign-in and auth shell, cloud/relay/managed tunnels and their relay test, hosted MCP apps (Poke, Claude, ChatGPT and others), the desktop relay-tunnel restart marker, the relay tracing defines, the credentials-changed account relist wakeup, and the T3 Account row.
• Analytics call sites that remained after fix: Ada no longer reports to upstream T3 services #245: the MCP tool analytics wrapper and handoff block in McpHttpServer, client analytics props and owners in ws.ts, the DelegatedTaskAnalytics and ProviderTurnAnalytics layers, turn analytics in RunExecutionService, and two analytics tests in core.test.
• Dead code: apps/mobile/src/state/auth.ts, which chore: delete dead code, unused deps, and duplicate helpers pingdotgg/t3code#9129 had deleted and the merge brought back.

gt-react/gt-react-native i18n uses bundled translations with no project id, so it makes no runtime network calls.

Custom UI (pingdotgg#35247)

Captures are in a secret gist. They come from the web client of vp run dev on seeded dev state, driven by a real Claude Haiku 5.5 turn calling t3_theme_set. Replay can't call MCP tools, so it couldn't drive this.

Before (default)	After the agent's theme	After the agent's revert

The prompt was "change the accent colour to teal (#0d9488) and add a small CSS tweak that gives the chat header a 2px solid bottom border in that teal". The agent published ada-teal (canvas #ffffff, accent #0d9488, CSS on [data-chat-header]) and switched to it. Asked to go back to the default, it set t3-code, which restored the default palette and the sidebar art.

• Surfaces. Web and desktop are supported. Mobile isn't, because upstream doesn't support it: mobile keeps its own themes and ignores environment themes.
• Ada naming. The tool descriptions, the agent theme guide, the theme eval cases and docs/user/appearance.md say Ada and ada theme set.
• Ada default and Pip. Publishing a theme never changes the default. The Pip logo shows under every theme. The sidebar art shows only under built-in themes, which is upstream design.
• Persistence. Agent themes are saved as files in the server's themes directory, and the default lives in server settings. Clients apply it once per generation, including after a reload or reconnect.

Conflicts with fork changes, and gaps fixed:

• The feature's own files merged cleanly. The fork-specific work was the rebrand of the model-facing text and docs/user/appearance.md.
• Revert to default was impossible. The standard look t3-code was not a selectable id, so agents could only switch to a built-in theme such as T3 Chat. t3-code can now be selected through the tool and ada theme set, but never published. The web client maps it back to the standard look and keeps the user's light/dark mode. Tests are in environmentTheme.test.ts.
• "Change the accent" turned Ada pink. On the first run (capture), Haiku put the accent only in colors. A colors-only theme starts from the stock pink palette, so Ada turned pink with a teal border. The guide now says to seed canvas and accent together, and that t3-code returns to the default look. On a rerun with the same prompt, Haiku seeded both and the whole palette followed. The new accent-change-seeds-canvas-and-accent eval case builds (--dry-run), but I didn't run it: this checkout has no Anthropic API key for evals.

Checks

• check:changed, run against origin/main with TMPDIR=/tmp and vite-plus Node 26. Format, lint, knip and typecheck failed on the first run. All of those failures came from the merge, and they're fixed in c330c3384e. On the rerun, format, lint (warnings only), knip and typecheck pass in every package touched, including evals. The test stage (876 files) hit the 900-second local timeout, so CI owns the full suite. Focused runs were green: about 1,700 server and 2,400 client tests during conflict work, plus the files touched afterwards.
• Fork guards: 31/31. Upstream migrations: 91 upstream plus 24 fork. check-removed-connect: green. Brand guard: 2/2. CI node:test scripts: green.
• Real client: one integrated test-t3-app pass on isolated, seeded dev state, covering the theme flow above.

Model: Claude Opus 5.5 with Claude Code subagents for the conflict slices.

🤖 Generated with Claude Code

[jayleaton]

Captain: accepting the agent-context growth (+386 B Claude system append, +50 B Codex context, +1.7 KB MCP tool list). It all comes from upstream 580b526's new tools and instructions, mainly the agent theme/CSS tool (pingdotgg#35247), which JJ asked for. The fork's tail-compat change must not add description text.

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	21.5 KiB	21.5 KiB	0 B (0.0%)	24.4 KiB	✅
Codex	Live turn messages	1	1	0 (0.0%)	8	✅
Claude	Total thread wire	5.0 KiB	5.0 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.8 KiB	3.8 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	21.9 KiB	21.9 KiB	+1 B (+0.0%)	24.4 KiB	✅
Claude	Live turn messages	1	1	0 (0.0%)	8	✅

Baseline: bfad67d · PR result: c330c33 · Source CI: failure

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.