[honzasladek]

The mobile app used a lot of cellular data. It subscribed to the PTY output of every terminal (also on the dashboard and in the background, where the Android connection service keeps the session open), and refetched the whole /v1/state snapshot uncompressed on every state_changed. A workspace with ~200 projects has a snapshot of ~235 KB.

Mobile (works against current servers)

• Stream only the terminal on screen. The transport gains a stream filter (RemoteClient::set_streamed_terminals) that survives reconnects. The app starts with nothing streamed and streams the selected terminal; none on the dashboard, in the background or after polling stops. A terminal that streams again gets a fresh snapshot.
• Coalesce state fetches. StateSyncPolicy: a minimum interval between fetches (the app uses 1 s; changes inside it become one fetch, the first change after a quiet window fetches at once), deferral while the app is in the background (one fetch on return), and no settings fetch, which the app never used. The desktop keeps the default policy.
• Data usage in Settings. The transport counts received payload bytes by kind (PTY, snapshots, state with the number of fetches, settings, other events). The state is fetched with transparent decompression off and decoded by hand, so the count is the size on the wire.
• Connection liveness counts any inbound frame instead of terminal output.

Server (needs a desktop release)

• gzip on the JSON routes (/v1/state, /v1/actions, usage, Claude session/transcript), per route so downloads and the WebSocket upgrade keep raw bodies. Not used over the local Unix socket. The snapshot above goes to ~32 KB.
• Git status no longer bumps state_version. It already has its own git_status_changed push, which the desktop and mobile clients patch into their state; the bump made every client refetch the whole snapshot on each git change. The web client now applies git_status_changed as well.
• set_system_stats: a client can opt out of the 2 s system_stats_changed push. The app does. Clients keeping the default never send it, so older servers are unaffected.
• terminal_attention in StateResponse: the last bell / OSC 9/777 notification time per terminal. Since the app no longer streams most terminals, it raises a terminal's unseen flag when its stamp moved between two states (no clock comparison; the first state is the baseline).

Compatibility

• New app, old server: streaming filter, coalescing and the data section work. No gzip; the old server answers set_system_stats with an invalid message error, which the app only logs; no attention stamps, so bells of terminals off screen are not flagged until the server is updated.
• Old app, new server: unchanged behaviour (gzip is negotiated, the new fields are optional).

Known trade-off

The output-based "running" dot for terminals that are not on screen is gone on mobile. Agent status covers agents; plain shells running a build no longer show as running.

Test plan

• cargo test for core, terminal, daemon-core, remote-server, transport, mobile-ffi, workspace, app-core, cli
• mobile: typecheck, lint, jest; web: tsc -b
• Transport connected to a running daemon over the local socket: one state fetch, no settings, no PTY bytes
• On a phone: switch terminals, background/foreground, check Settings → Data over a few hours
• Desktop: git badges still update; host metrics still shown for remote connections

🤖 Generated with Claude Code