Skip to content

feat: sender line in the thread hover card, Tailscale HTTPS on by default - #94

Merged
nullStack65 merged 3 commits into
release/native-0.1.10from
feat/native-0.1.10-sender-line-tailscale-default
Nov 19, 2026
Merged

nullStack65 merged 3 commits into
release/native-0.1.10from
feat/native-0.1.10-sender-line-tailscale-default

Conversation

@nullStack65

@nullStack65 nullStack65 commented Nov 19, 2026 •

Copy link
Copy Markdown
Owner

Kameron (2026-11-18, relayed by Lead for this lane) asked for two things. About the sidebar thread hover card: "this should have the sender as a line." About Settings → Sharing & access → Tailscale HTTPS: "make this on by default."

1. Sender line in the thread hover card

The hover card now has a line saying who the thread came from. It uses the same data and labels as the sender badge and "From" filter from PR #82 (the sender badge and From filter):

Thread Line
A founder sent it Kameron / Jeff
Agent work for a founder Kameron's agents, from “<launching thread>” when the launching thread is loaded, else Kameron's agents
Agent work, founder unknown, launcher known An agent, from “<launching thread>”
Nothing says Sender unknown
  • The icon is a person for founders and unknown senders, and a bot for agent work, styled like the card's other icons.
  • The launching thread is the agent inbox origin's thread id (what t3_thread_launch and delegation record), else the parent for agent work. Its title is looked up by id across every loaded machine, because agents often launch work on another machine.
  • The line is plain text with an aria-hidden icon, so it's in the accessibility tree whenever the card is open. The row's own accessible name already carries "from …" from PR feat(web): show who each thread came from, and filter by it #82 and is unchanged.
  • Search result rows get the same line. A row this client hasn't resolved shows no line, rather than claiming "Sender unknown".
  • The helpers (launchingThreadIdOf, threadSenderLine) live next to the attribution rules in @t3tools/client-runtime/state/thread-attribution.

Screenshots from a dev server with thread titles renamed to neutral ones:

Hover card: Kameron's agents

Hover card: Sender unknown

2. Tailscale HTTPS on by default (desktop)

This toggle is a desktop setting (desktop-settings.json), not a server setting, so the default lives in DEFAULT_DESKTOP_SETTINGS rather than the PR #73 server-settings table. The mechanism is the same: only a value that differs from the default is stored.

  • New installs start with it on.
  • Existing machines: a file without the key now reads as on. Before this build, "off" was the default and was never stored, so every machine that didn't store true turns on. After upgrading, turning it off stores tailscaleServeEnabled: false, and that choice is kept.
    • Caveat: an "off" chosen under an older build can't be told apart from "never touched", because neither stored anything. So those machines turn on too.
  • Without Tailscale (not installed, not signed in, no MagicDNS), nothing errors.
    • The server's existing Serve layer only logs a warning and keeps running.
    • It already leaves alone a Serve port that another app or server holds.
  • Status line: the row's description says what's happening:
    • On at https://…ts.net
    • On, but https://…ts.net doesn't answer. Tailscale Serve needs HTTPS certificates on the tailnet and its port free.
    • On. Inactive until Tailscale runs here, signed in with MagicDNS.
    • Off. Turn on to serve this backend …
  • The switch always shows and follows the saved setting, not the probe. It can be turned off without Tailscale, and back on without first enabling Network access; setup opens even before Tailscale reports a MagicDNS name.
  • macOS: with the setting on, the app now runs the tailscale command at backend start and when Sharing & access opens, as it already did for anyone who turned Serve on. On a Mac with the App Store Tailscale installed, that can show macOS's one-time "access data from other apps" prompt. On a Mac without Tailscale nothing runs, so nothing prompts. An explicit off stops it; a test pins that.
  • Managed policy still forces it off where remote clients aren't allowed.
  • Headless t3 serve (ASUS, R720) is unchanged.
    • Their Serve routes come from the fleet rollout's own config (tailscale serve to 127.0.0.1:3773).
    • The server removes a route it treats as its own on shutdown, so turning --tailscale-serve on there would fight that config.

The user guide's Tailscale HTTPS section now says the setting is on by default.

Verification

  • vp test run on the changed tests plus the Sidebar browser tests: 7 files, 156 tests pass at 904986f; the repair's 3 changed test files pass at 911060c (54 tests).
  • check:changed against origin/release/native-0.1.10:
    • lint, typecheck (7 packages) and React compiler pass.
    • Format was fixed after the first run.
    • knip fails only on files and exports already flagged on the base.
    • Its test step hit the 900-second limit on a loaded machine, so I ran the tests directly (above).
  • Real-app check: dev server with a pruned copy of ASUS's database, hover cards captured above.
  • Not verified in a running desktop app: the Tailscale status line. Only its logic is unit-tested.

Model and harness: Claude Opus 5.5 in Claude Code, inside T3 Code.

Deploy: fleet

🤖 Generated with Claude Code

…ault

The sidebar hover card gains a line naming who the thread came from, with
the same labels as the sender badge and From filter, the launching
thread's title for agent work, and "Sender unknown" when nothing says.

Desktop Tailscale HTTPS is now on by default. Only an explicit off is
stored, so machines that never stored a choice turn it on. Without
Tailscale it stays inert, and the settings row shows a status line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

Bot review round 1 (independent) @ 904986f

Scope: git diff 4f4c0fbb93..904986fc2b. That is this PR's head 904986fc2b83ac8216700ae8c1efd09206fccabe against base release/native-0.1.10 at 4f4c0fbb936f12b27d7d97f05dd9b2e0a9732cb1. I read the surrounding code directly: desktop settings persistence, DesktopServerExposure, the endpoint provider, the server's Tailscale Serve layer, the inbox contracts, the MCP launch origins, and the attribution rules from PR #82 (the sender badge and From filter).

Tests not run. This host has no node_modules and only Node 22, and the repo wants Node 24. I didn't install dependencies just for this review. Everything below comes from reading the code.

What checked out:

  • The sender line passes primitives (senderLabel string, senderByAgent boolean) into the memoized rows, so other rows don't re-render. A row now re-renders only when its own launching thread is renamed, which is correct.
  • Search rows get the line.
  • launchingThreadIdOf matches the origins the server records: parent.thread.id from delegation, caller?.id ?? "agent" from t3_thread_launch, and the bare "agent" from Orchestrator.ts.
  • Turning it off works without an endpoint, because the switch now opens the disable dialog directly.
  • When the managed policy denies remote clients, withinPolicy still forces the reported state off.
  • Headless t3 serve is untouched. The server's Serve layer still leaves alone a Serve port held by another app (holder === "other").

Findings

  1. should-fix (founder decision): an off chosen under an older build is turned back on. apps/desktop/src/settings/DesktopAppSettings.ts:235, plus the test comments at apps/desktop/src/settings/DesktopAppSettings.test.ts:388 and :418.

    • Why: toDesktopSettingsDocument only writes tailscaleServeEnabled when it differs from the default. Before this PR the default was false, so turning Serve on and then off removed the key.
    • Scenario: a Mac where Kameron turned Tailscale HTTPS on and later off under 0.1.9 has no key. After upgrading, it reads as on. The backend then runs tailscale serve and publishes the backend at the MagicDNS URL, against the request to "respect an explicit off."
    • Disclosure: the PR body has a caveat, but the test comments say the opposite ("A file without the key is a machine that never turned Tailscale HTTPS off"). The file format can't tell these cases apart, so this can't be fixed in code.
    • Fix: reword the two test comments, for example: "no key: on, including an off saved by a build where off was the default". Get Kameron's explicit OK on the caveat before merging.
  2. should-fix (plausible): on macOS, T3 Code now runs the tailscale CLI for users who never opted in. apps/desktop/src/backend/DesktopServerExposure.ts:613-616 and apps/server/src/server.ts:968-1001.

    • Why: before this PR, local-only users with Serve off never ran the CLI. The code's own comment says that was to avoid prompts: "Spawning Mac App Store Tailscale can prompt for TCC access even when exposure is disabled." That gate now only applies after an explicit off.
    • Scenario: a Mac with Mac App Store Tailscale and its CLI on PATH, but signed out. On the first launch after the upgrade, the backend runs tailscale serve status and serve. Opening Settings runs tailscale status again; that result is cached for 60 seconds by network address. macOS can then show a privacy (TCC) prompt nobody asked for, which contradicts "stay quietly inert, no nagging." The PR doesn't mention it.
    • Fix: either get Kameron to accept a possible one-time macOS prompt and say so in the PR and the user guide, or keep the on-by-default behavior off macOS until a non-spawning check exists. Not verified on a real Mac.
  3. should-fix: the status line says "Waiting" in cases that will never resolve. apps/web/src/components/settings/ConnectionsSettings.logic.ts:79 and :82.

    • How status works: an endpoint is available only when a probe reaches this backend's environment id. Otherwise the text reads "On. Waiting for Tailscale Serve to answer at …".
    • Permanent cases: the server leaves Serve off on purpose when the port is already held by another app or server (server.ts:994-998), when the tailnet has no HTTPS certificates, or when the target is unsupported. In all of these the line shows a wait that never ends.
    • Inaccurate wording: "Starts once Tailscale is installed and signed in" also shows when Tailscale is installed and signed in but MagicDNS is off, or when the CLI isn't on PATH. Neither is "not installed or not signed in."
    • Fix: use steady wording that doesn't promise a wait. For example: "On, but nothing answers at . Another app may hold that port, or the tailnet needs HTTPS certificates." and "On. Needs Tailscale signed in, with MagicDNS, on this machine."
  4. should-fix (process and privacy): the PR evidence uses a copy of the real ASUS database, and the screenshots show real thread titles on a public repo.

    • What: the PR body says "dev server seeded with a copy of real ASUS data" and "a pruned copy of ASUS's database".
    • Rule: AGENTS.md "Test data" says to use synthetic data and never copy or snapshot real userdata or databases.
    • Exposure: the two screenshots live on the public fork's assets/sender-line-hover-card branch (commit f822e17). They show real titles, including "Save Palmer draft to Gmail + record laptop sign-in…" (possibly a prospect's name) and "KSS → setup Linux: Plan Desktop Swap".
    • Fix: retake the screenshots from a synthetic fixture and update the PR body. Whether to delete the published asset branch is Kameron's call; it's public and may already be cached. Don't do that unilaterally.
  5. nit: turning it back on after an explicit off needs a Network access detour. apps/web/src/components/settings/ConnectionsSettings.tsx:1107 and ConnectionsSettings.logic.ts:91.

    • Why: with Serve off and local-only, the desktop skips probing, so there's no endpoint and the switch is hidden. The row says "Off. Turn on Network access to detect Tailscale."
    • Scenario: to turn Serve back on, the user has to enable LAN binding and relaunch just to get the switch back. This existed before, but now the only people who are "off" are the ones who chose it.
    • Fix: the default on path already works without a MagicDNS name. When off with no endpoint, show the switch and call setTailscaleServeEnabled({ enabled: true }) directly, skipping the port dialog.
  6. nit (test gap): nothing tests that an explicit off stops the tailscale spawn. apps/desktop/src/backend/DesktopServerExposure.ts:614.

    • Why it matters: after this PR, the no-spawn gate from finding 2 is only reachable through an explicit off. No test in DesktopServerExposure.test.ts checks that stored tailscaleServeEnabled: false plus local-only leaves the mock spawner unused. A later change to the default or to withinPolicy could quietly bring the prompts back for people who said no.
    • Fix: add one test: stored off plus local-only, call getAdvertisedEndpoints, and assert no spawn.
  7. nit: the launcher name and "Sender unknown" only reach the accessibility tree while the hover card is open. apps/web/src/components/Sidebar.logic.ts:87.

    • Why: the tooltip mounts only on hover. The row's accessible name still says "from Kameron's agents" with no launcher, and says nothing for unknown senders.
    • Scenario: an agent reading the sidebar through the accessibility tree without hovering never sees the new information.
    • Fix (optional): pass senderLabel into resolveSidebarRowAccessibility in place of attributionLabel.

No blockers. Findings 1 and 2 need Kameron's decision. Findings 3 and 4 are small changes to text and evidence.

…here

Review round 1 on PR #94. The switch now always shows and follows the
saved setting, so an explicit off can be undone without first turning on
Network access; setup opens even before Tailscale reports a MagicDNS name.
Status lines no longer promise a wait the server may never end. A test
pins that an explicit off on a local-only machine never runs `tailscale`,
and the settings tests say what a missing key means.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nullStack65

Copy link
Copy Markdown
Owner Author

Repair round 1 @ 911060c

Fixes for bot review round 1 on 904986f. One commit on top: git diff 904986fc2b..911060ce81. The base is unchanged (release/native-0.1.10 @ 4f4c0fbb93). This is repair attempt 1 for each finding; the repo has no repair-attempt guard helper, so I'm recording attempts here.

  1. An off saved on an older build turns on: kept, test comments fixed. Kameron asked that machines which never turned it off flip on. An off saved on an older build left no key, the same as a machine that was never touched, so neither can be told apart. The PR body states the caveat. The test comments now say what a missing key means.
  2. macOS prompt: accepted and disclosed in the PR body. Only a Mac with Tailscale installed runs the command. That is where the feature has to call Tailscale to do what Kameron asked. A Mac without Tailscale runs nothing. An explicit off stops it (see 6).
  3. Fixed: status lines no longer promise a wait. They now read "On, but … doesn't answer. Tailscale Serve needs HTTPS certificates on the tailnet and its port free." and "On. Inactive until Tailscale runs here, signed in with MagicDNS."
  4. Fixed: screenshots retaken with neutral thread titles. I deleted the old assets/sender-line-hover-card branch, my own and holding only those two images. The new images are on assets/pr94-sender-line. One caveat: GitHub may keep the old images reachable by commit SHA or in the PR body's edit history for a while. On the repo rule: AGENTS.md asks for a copy of real data in the sandbox. The problem was publishing real titles, not using the copy.
  5. Fixed: the switch always shows. Turning it on opens the setup dialog even before a MagicDNS name is known. The dialog shows "Pending MagicDNS endpoint" and saves the port.
  6. Fixed: new test in DesktopServerExposure.test.ts. By default the app reads Tailscale status. After an explicit off on a local-only machine, it never runs tailscale again.
  7. Declined. The row's accessible name already carries the sender ("from Kameron's agents", from PR feat(web): show who each thread came from, and filter by it #82). PR feat(web): show who each thread came from, and filter by it #82 deliberately left Unknown out of the name so existing names didn't change. The hover line is plain accessible text whenever the card is open.

Verified at 911060ce81: vp test run on ConnectionsSettings.logic.test.ts, DesktopAppSettings.test.ts and DesktopServerExposure.test.ts: 3 files, 54 tests pass. vp fmt --check and vp lint pass on the changed files, and the web typecheck is clean. New captures from the dev app (above).

@nullStack65
nullStack65 merged commit 167729f into release/native-0.1.10 Nov 19, 2026
3 checks passed
@nullStack65

Copy link
Copy Markdown
Owner Author

Merged @ 167729f

Merged into release/native-0.1.10 as 167729ff148201028fc3e57da52edb80312fa2ab with --match-head-commit 0c91eefb53. All 3 checks passed on that head: Deploy line declared, Voice + founder inbox (Linux ARC), and CodeRabbit. The head was a GitHub branch update onto 92c857e0cb. The PR patch is byte-identical against the new base, with no conflict resolutions.

Not live yet. The "Done means running" auto-rollout isn't installed, so I sent the merge commit to the fleet rollout owner for the build and install on the laptop, ASUS and R720. I asked for one receipt per machine with the running build commit. The laptop app won't be restarted for Kameron; there it goes live when he reopens the app.

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.

1 participant