[httpsphl]

Last of three slices of M4: messages, deliveries and tasks in the app. This PR is built on #6, which is built on #5. Merge them in order (merge commits); this diff then shrinks to the two commits below. With it, M4 is complete: crews, bots, terminals, the timeline and failed deliveries all work from the app, without opening a terminal.

What changes

• Store (spec 15.1):
  • Deliveries by message: the 500 most recently updated plus every dead one, then each delivery.changed.
  • Every task, then each task.changed.
  • Loading moves to store/sync.ts.
  • Messages are not stored globally: each timeline pages its own.
• Timelines:
  • The crew timeline (Timeline tab) and a bot's (Messages tab: messages to or from it).
  • Oldest first. The view follows new messages while the owner is at the end, and older pages load on demand.
  • Each message shows sender, recipient, a task or result tag with the task's live state, and its delivery: waiting, retrying (next try and last error), delivering, delivered, or not delivered with Retry.
• Composer:
  • The owner writes to any bot of the crew from the crew timeline, or to the bot itself from its Messages tab.
  • Ctrl+Enter sends, with the daemon's length limit shown.
• Failed deliveries:
  • A title-bar button appears while any message to an active bot gave up.
  • It opens the list, with each message, its error and Retry, plus Retry all.
  • Dead deliveries to archived bots are left out, because they were abandoned on purpose.
• Tasks tab:
  • Open tasks by default, every task on demand.
  • Each task shows status (color, icon and label), requester and assignee, its hop, the deadline or when it settled, and the result.
• Taskbar overlay (spec 15.2):
  • A dot on the app icon while a bot waits for approval or sign-in, or a message to an active bot was not delivered.
  • It uses the window's setOverlayIcon (new permission core:window:allow-set-overlay-icon), with no command of our own.
• Test support: FakeConversation plays the bots and the courier in tests (say, deliver, task), and src/test/app.tsx renders the whole app. The fake's clock is now the real one.

Found in the real test (spec 9.3)

Claude Code 2.1.284 presents everything from the inbox as coming from another session and not typed by the user, including the owner's message. It also adds a note that a peer cannot grant permissions or approve a pending prompt.

So a timeline message has a colleague's authority. To approve a prompt or change settings as the bot's user, the owner types in the bot's terminal. The "Needs approval" notice already points there.

Tests

• messages.test.tsx:
  • the crew timeline with a bot's message and its delivery;
  • the owner sending to a chosen recipient: waiting, then delivered when the courier moves;
  • the bot tab showing only its messages, with a fixed recipient and Ctrl+Enter;
  • paging 60 messages;
  • a dead delivery retried from the title bar ("Retry all", with the message body shown) and from the timeline;
  • the taskbar mark following a bot in needs_approval.
• Tasks: the open filter, hops and deadline, the "All" filter with results, and a task changing live.
• Totals: 41 app tests and 140 Rust tests. pnpm check, pnpm build, clippy (including the Tauri capability check) and fmt pass. Each commit passes on its own.

Checked with real Claude Code 2.1.284

The test used the dev preview connected to a real dev daemon.

1. A message was sent from the bot's Messages tab.
2. The courier delivered it, and the UI showed Delivered.
3. The bot went Working, then Idle, and answered "PONG" in its terminal, below the [botloft] from the owner · crew … envelope.

Manual test (Tauri window)

1. With a bot waiting on a permission prompt, the taskbar icon shows the dot, and it clears once the prompt is answered.
2. A dead delivery is hard to cause by hand: a paused or stopped bot only defers its messages, which do not count as attempts. The title-bar list and Retry are covered by the component tests.