[httpsphl]

Bots now run as headless Claude Code and the app talks to them in a chat, like a messaging app, instead of showing a terminal. The design is in ADR 0001 and spec v0.2 (sections 7 to 10, 15, 17 M4.1 and 19).

Daemon

• Each bot runs claude -p with stream-json on stdin and stdout; no PTY, hooks or named-pipe inbox. portable-pty, the hooks and the inbox code are gone.
• The daemon turns what the bot prints into chat items: the owner's and other bots' messages, replies, tool calls, approvals, turn ends and notices. Replies are also sent as live text while the bot writes them. Protocol 2 adds chat.history, approvals.answer, attachments.read, chat.item and chat.delta; terminal.* is gone.
• The courier writes messages into stdin. The replayed user message marks a delivery as read. What a process never read goes back to the queue when the process ends.
• Tool permissions go through the permission_prompt MCP tool and wait for the owner (60 min by default). A request expires when its connection drops, when the process ends or when the daemon starts.
• Bot state comes from the stream: busy, needs_approval, rate_limited (with the account usage in system.status) and auth_error. Sessions resume after a restart, and a resume that dies at once starts fresh.
• Owner messages can carry up to 10 files, saved in the bot's folder; images also go inline.

App

• Bot chat: the owner's bubbles on the right with delivered and read receipts; other bots on the left.
  • Replies render as markdown and stream in. Links open in the default browser through a new open_url command (http and https only); images in replies are not fetched.
  • Tool calls show as compact lines that open to show input and output. Approvals show as cards with Allow, Deny and a note.
• Composer: Enter sends and Shift+Enter breaks the line. Files can be picked, pasted or dropped, and images show as thumbnails, read back after a reload.
• The sidebar lists bots as conversations with their state and last activity. Bot details moved to a side panel. xterm.js is gone.

Checks

cargo fmt --check, cargo clippy --workspace --all-targets -D warnings, cargo test --workspace and pnpm check pass locally.

New tests:

• daemon: tests/chat.rs (items, deltas, approvals allowed, denied, timed out, unknown and expired on exit), tests/supervisor_turns.rs, and tests/attachments.rs; courier and supervisor tests were rewritten for stdin.
• app: chat.test.tsx and rows.test.ts.

Tested with real Claude Code (2.1.284)

Done on a dev daemon with a fresh BOTLOFT_HOME and the app preview (pnpm dev, ?live=<port>). Results are in spec section 19. To repeat:

1. Start a dev daemon with its own BOTLOFT_HOME, open the app against it, create a crew and a bot.
2. Ask the bot to create a file. Expected:
   • the message shows "Read";
   • the reply streams in;
   • a Write line appears with an approval card, and the bot shows "Needs approval".
3. Allow, and the turn ends with "Done in …". Then do it again, waiting more than 60 s before allowing: it still goes through.
4. Attach an image and ask about it, then ask for something that needs a command and Deny it with a note. The bot describes the image, says it did not run the command, and mentions the note.
5. Reload the page: the image thumbnail is still there.
6. Restart the bot, and separately restart the daemon. Asking about earlier messages shows the bot resumed its session.
7. Create a second bot and ask the first to send_message it. The note shows in the second bot's chat, and its answer reaches the first.
8. Stop the daemon: no claude.exe with mcp__botloft__permission_prompt in its command line remains.