feat(agents): durable tmux-backed terminal sessions + fit/chrome fixes

- Run agent sessions under an isolated tmux server (its own systemd scope, with a
  keepalive session) so they survive a castle-api restart and stay rediscoverable;
  each WebSocket is a tmux attach client. Falls back to the in-memory pty backend
  when tmux is absent or CASTLE_API_AGENT_BACKEND=memory.
- Refactor the session layer behind one Attachment interface (create/list/attach/
  close), keeping resume-by-id, agent-native history, and continue modes.
- Fix the terminal not refilling when the dock is maximized (explicit refit signal
  + instant size transition so xterm measures the final box).
- On narrow screens, go full-bleed and strip chrome (header/border/padding) when
  maximized.

app/src/components/AgentTerminal.tsx

@@ -4,13 +4,14 @@ import { Terminal } from "@xterm/xterm"
 import { FitAddon } from "@xterm/addon-fit"
 import "@xterm/xterm/css/xterm.css"
 import { apiClient } from "@/services/api/client"
+import { cn } from "@/lib/utils"
 
 type Status = "connecting" | "connected" | "exited" | "closed" | "error"
 
 interface AgentTerminalProps {
   // Agent name to launch, or "terminal" for the plain login shell.
   agent: string
-  // When set, resume this existing session instead of creating a new one.
+  // When set, resume this existing live session (replaying its scrollback).
   resumeId?: string
   // Fired once the backend confirms the session id (new or resumed).
   onSession?: (id: string) => void
@@ -20,6 +21,13 @@ interface AgentTerminalProps {
   mode?: "continue"
   // Resume one of the agent's OWN past sessions by its id (agent-native).
   resumeSessionId?: string
+  // Bump this whenever the terminal's container is resized by an ancestor (e.g.
+  // the dock expanding). Triggers a refit that ResizeObserver alone can miss
+  // when the size change is driven by a class swap rather than layout reflow.
+  fitSignal?: number
+  // Strip chrome (header, border, padding) on narrow screens — used when the
+  // dock is maximized so the terminal goes edge-to-edge on a phone.
+  compact?: boolean
 }
 
 export function AgentTerminal({
@@ -29,12 +37,32 @@ export function AgentTerminal({
   fill,
   mode,
   resumeSessionId,
+  fitSignal,
+  compact,
 }: AgentTerminalProps) {
   const containerRef = useRef<HTMLDivElement>(null)
+  const termRef = useRef<Terminal | null>(null)
+  const fitRef = useRef<FitAddon | null>(null)
+  const wsRef = useRef<WebSocket | null>(null)
   const [status, setStatus] = useState<Status>("connecting")
   const [detail, setDetail] = useState<string>("")
   const [maximized, setMaximized] = useState(false)
 
+  // Fit to the container, then tell the backend pty the new size.
+  const refit = () => {
+    const term = termRef.current
+    const ws = wsRef.current
+    if (!term) return
+    try {
+      fitRef.current?.fit()
+    } catch {
+      /* container not laid out yet */
+    }
+    if (ws && ws.readyState === WebSocket.OPEN) {
+      ws.send(JSON.stringify({ type: "resize", cols: term.cols, rows: term.rows }))
+    }
+  }
+
   useEffect(() => {
     const term = new Terminal({
       convertEol: false,
@@ -46,6 +74,8 @@ export function AgentTerminal({
     const fit = new FitAddon()
     term.loadAddon(fit)
     term.open(containerRef.current!)
+    termRef.current = term
+    fitRef.current = fit
     try {
       fit.fit()
     } catch {
@@ -61,17 +91,12 @@ export function AgentTerminal({
           : `/agents/${agent}/session`
     const ws = new WebSocket(apiClient.wsUrl(path))
     ws.binaryType = "arraybuffer"
+    wsRef.current = ws
     const enc = new TextEncoder()
 
-    const sendResize = () => {
-      if (ws.readyState === WebSocket.OPEN) {
-        ws.send(JSON.stringify({ type: "resize", cols: term.cols, rows: term.rows }))
-      }
-    }
-
     ws.onopen = () => {
       setStatus("connected")
-      sendResize()
+      refit()
       // On resume, full-screen TUIs (opencode, claude) don't know a new client
       // attached and won't repaint. A size change forces SIGWINCH → a full
       // redraw of the current frame, so the prior session becomes visible.
@@ -81,7 +106,7 @@ export function AgentTerminal({
           ws.send(
             JSON.stringify({ type: "resize", cols: Math.max(1, term.cols - 1), rows: term.rows }),
           )
-          setTimeout(sendResize, 60)
+          setTimeout(refit, 60)
         }, 80)
       }
       term.focus()
@@ -116,14 +141,7 @@ export function AgentTerminal({
       if (ws.readyState === WebSocket.OPEN) ws.send(enc.encode(d))
     })
 
-    const ro = new ResizeObserver(() => {
+    const ro = new ResizeObserver(() => refit())
-      try {
-        fit.fit()
-      } catch {
-        /* ignore */
-      }
-      sendResize()
-    })
     if (containerRef.current) ro.observe(containerRef.current)
 
     return () => {
@@ -131,10 +149,27 @@ export function AgentTerminal({
       dataDisp.dispose()
       ws.close()
       term.dispose()
+      termRef.current = null
+      fitRef.current = null
+      wsRef.current = null
     }
     // Remounting (new agent / resume target / mode) tears this down and restarts.
   }, [agent, resumeId, mode, resumeSessionId, onSession])
 
+  // An ancestor changed our size (dock expanded/restored/reopened). ResizeObserver
+  // usually catches it, but a class-swap + CSS can settle a frame late, so refit
+  // across a few frames until the new size stabilizes.
+  useEffect(() => {
+    let raf = 0
+    const start = performance.now()
+    const tick = () => {
+      refit()
+      if (performance.now() - start < 400) raf = requestAnimationFrame(tick)
+    }
+    raf = requestAnimationFrame(tick)
+    return () => cancelAnimationFrame(raf)
+  }, [fitSignal])
+
   const statusColor =
     status === "connected"
       ? "text-green-400"
@@ -149,15 +184,21 @@ export function AgentTerminal({
 
   return (
     <div
-      className={
+      className={cn(
         maximized
           ? "fixed inset-0 z-50 bg-[var(--background)] flex flex-col"
           : fill
             ? "h-full flex flex-col bg-black/40 border border-[var(--border)] rounded-lg overflow-hidden"
-            : "bg-black/40 border border-[var(--border)] rounded-lg overflow-hidden"
+            : "bg-black/40 border border-[var(--border)] rounded-lg overflow-hidden",
-      }
+        compact && "max-sm:border-0 max-sm:rounded-none",
+      )}
     >
-      <div className="flex items-center justify-between px-3 py-1.5 border-b border-[var(--border)] text-xs text-[var(--muted)]">
+      <div
+        className={cn(
+          "flex items-center justify-between px-3 py-1.5 border-b border-[var(--border)] text-xs text-[var(--muted)]",
+          compact && "max-sm:hidden",
+        )}
+      >
         <span className="flex items-center gap-2">
           <span>terminal: {agent}</span>
           <span className={statusColor}>
@@ -180,7 +221,10 @@ export function AgentTerminal({
       </div>
       <div
         ref={containerRef}
-        className={maximized || fill ? "flex-1 min-h-0 p-2" : "h-[560px] p-2"}
+        className={cn(
+          maximized || fill ? "flex-1 min-h-0 p-2" : "h-[560px] p-2",
+          compact && "max-sm:p-0",
+        )}
       />
     </div>
   )

app/src/components/AssistantDock.tsx

@@ -99,12 +99,15 @@ export function AssistantDock() {
           terminal's WebSocket survives minimize. */}
       <div
         className={cn(
-          "fixed z-40 flex flex-col",
+          "fixed z-40 flex flex-col bg-[var(--background)] shadow-2xl",
           expanded
-            ? "inset-4"
+            ? // Maximized: full-bleed on narrow screens (no margin/border/radius),
-            : "bottom-4 right-4 w-[760px] max-w-[calc(100vw-2rem)] h-[600px] max-h-[calc(100vh-2rem)]",
+              // ins-4 with chrome on sm+.
-          "bg-[var(--background)] border border-[var(--border)] rounded-xl shadow-2xl",
+              "inset-0 rounded-none border-0 sm:inset-4 sm:rounded-xl sm:border sm:border-[var(--border)]"
-          "transition-all duration-200",
+            : "bottom-4 right-4 w-[760px] max-w-[calc(100vw-2rem)] h-[600px] max-h-[calc(100vh-2rem)] rounded-xl border border-[var(--border)]",
+          // Animate only open/close (opacity + translate); size changes (expand)
+          // are instant so the terminal's fit measures the final box immediately.
+          "transition-[opacity,transform] duration-200",
           open
             ? "opacity-100 translate-y-0"
             : "opacity-0 translate-y-[calc(100%+2rem)] pointer-events-none",
@@ -114,7 +117,14 @@ export function AssistantDock() {
         {/* Header */}
         <div className="flex items-center gap-2 px-3 py-2 border-b border-[var(--border)]">
           <Bot size={16} className="text-[var(--primary)] shrink-0" />
-          <div className="flex flex-wrap items-center gap-1.5 min-w-0">
+          <div
+            className={cn(
+              "flex flex-wrap items-center gap-1.5 min-w-0",
+              // Maximized on a narrow screen: drop the launcher pills for space
+              // (shrink to switch agents). Visible again at sm+.
+              expanded && "max-sm:hidden",
+            )}
+          >
             <button
               onClick={() => launch(SHELL)}
               className={cn(
@@ -190,7 +200,7 @@ export function AssistantDock() {
         </div>
 
         {/* Body */}
-        <div className="relative flex-1 min-h-0 p-2">
+        <div className={cn("relative flex-1 min-h-0 p-2", expanded && "max-sm:p-0")}>
           {active ? (
             <AgentTerminal
               key={active.key}
@@ -200,6 +210,8 @@ export function AssistantDock() {
               resumeSessionId={active.resumeSessionId}
               onSession={onSession}
               fill
+              compact={expanded}
+              fitSignal={(open ? 1 : 0) + (expanded ? 2 : 0)}
             />
           ) : (
             <div className="h-full flex items-center justify-center text-sm text-[var(--muted)]">

castle-api/src/castle_api/agent_sessions.py

@@ -1,49 +1,75 @@
 """Live agent/terminal session registry.
 
-A session is a running PTY (an agent CLI or a shell) whose lifetime is decoupled
+A session is a running terminal (an agent CLI or a shell) whose lifetime is
-from any WebSocket connection: closing the browser tab detaches, it does not
+decoupled from any WebSocket connection — closing a tab detaches, it does not
 kill the process. Sessions can be listed, re-attached (resumed) by id, and
-explicitly terminated. Each session keeps a bounded scrollback buffer so a
+explicitly terminated.
-re-attaching client can repaint the terminal.
 
-This is deliberately in-memory and single-node: sessions do not survive a
+Two backends behind one interface, chosen at startup:
-castle-api restart. That's the right scope for a personal dashboard — for true
+
-cross-restart persistence you'd launch the agent under tmux, which breaks the
+- **tmux** (default when `tmux` is on PATH): each session is a tmux session, so
-"castle just runs a command" agnosticism.
+  it **survives a castle-api restart** and is rediscoverable. The tmux server is
+  started inside its own systemd scope so it isn't in castle-api's cgroup (which
+  systemd would kill on restart); tmux's own systemd integration puts each pane
+  in its own scope too. Each WebSocket is a `tmux attach` client, so multi-client
+  and repaint-on-attach come for free.
+- **memory** (fallback when tmux is absent, or `CASTLE_API_AGENT_BACKEND=memory`):
+  castle-api owns the pty directly and fans output out to subscribers with a
+  scrollback buffer. Simpler, but sessions die on restart.
+
+The WebSocket handler is backend-agnostic: it calls `manager.create/get_info/
+list/close`, then `manager.attach(id)` and drives the returned `Attachment`.
 """
 
 from __future__ import annotations
 
 import asyncio
+import logging
+import os
+import shlex
+import shutil
 import time
 import uuid
-from dataclasses import dataclass, field
+from collections.abc import AsyncIterator
-from pathlib import Path
+from typing import Protocol
+
+from castle_core.config import CASTLE_HOME
 
 from castle_api.pty_session import PtySession
 
-# Cap the per-session replay buffer. Raw terminal bytes; enough to repaint a
+logger = logging.getLogger(__name__)
-# full-screen TUI and some history without unbounded growth.
+
-_SCROLLBACK_BYTES = 256 * 1024
+_SCROLLBACK_BYTES = 256 * 1024  # memory backend replay cap
-# Reap sessions that have exited and sat idle longer than this (seconds).
+_EXITED_TTL = 3600  # memory backend: reap exited sessions after this
-_EXITED_TTL = 3600
 
 
-@dataclass
+class Attachment(Protocol):
-class AgentSession:
+    """One WebSocket's live view of a session."""
-    """One live PTY plus its scrollback and attached subscribers."""
 
-    id: str
+    def reader(self) -> AsyncIterator[bytes]: ...
-    agent: str  # agent name, or "terminal" for the plain shell
+    def write(self, data: bytes) -> None: ...
-    command: str
+    def resize(self, cols: int, rows: int) -> None: ...
-    cwd: str
+    async def aclose(self) -> None: ...
-    created_at: float
+
-    pty: PtySession | None = None
+
-    scrollback: bytearray = field(default_factory=bytearray)
+# ---------------------------------------------------------------------------
-    subscribers: set[asyncio.Queue[bytes | None]] = field(default_factory=set)
+# In-memory backend
-    exited: bool = False
+# ---------------------------------------------------------------------------
-    exit_code: int | None = None
+
-    exited_at: float | None = None
+
+class _MemSession:
+    def __init__(self, sid: str, agent: str, command: str, cwd: str) -> None:
+        self.id = sid
+        self.agent = agent
+        self.command = command
+        self.cwd = cwd
+        self.created_at = time.time()
+        self.pty: PtySession | None = None
+        self.scrollback = bytearray()
+        self.subscribers: set[asyncio.Queue[bytes | None]] = set()
+        self.exited = False
+        self.exit_code: int | None = None
+        self.exited_at: float | None = None
 
     def _on_output(self, data: bytes) -> None:
         self.scrollback.extend(data)
@@ -53,7 +79,6 @@ class AgentSession:
             try:
                 q.put_nowait(data)
             except asyncio.QueueFull:
-                # Slow client: drop it; the WS side will notice on its next send.
                 self.subscribers.discard(q)
 
     def _on_exit(self) -> None:
@@ -61,15 +86,7 @@ class AgentSession:
         self.exit_code = self.pty.proc.returncode if self.pty else None
         self.exited_at = time.time()
         for q in list(self.subscribers):
-            q.put_nowait(None)  # EOF sentinel
+            q.put_nowait(None)
-
-    def subscribe(self) -> asyncio.Queue[bytes | None]:
-        q: asyncio.Queue[bytes | None] = asyncio.Queue(maxsize=256)
-        self.subscribers.add(q)
-        return q
-
-    def unsubscribe(self, q: asyncio.Queue[bytes | None]) -> None:
-        self.subscribers.discard(q)
 
     @property
     def running(self) -> bool:
@@ -91,58 +108,78 @@ class AgentSession:
         }
 
 
-class SessionManager:
+class _MemAttachment:
-    """In-memory registry of live agent sessions."""
+    def __init__(
+        self, session: _MemSession, q: asyncio.Queue[bytes | None], snap: bytes
+    ):
+        self._s = session
+        self._q = q
+        self._snap = snap
+
+    async def reader(self) -> AsyncIterator[bytes]:
+        if self._snap:
+            yield self._snap
+        while (chunk := await self._q.get()) is not None:
+            yield chunk
+
+    def write(self, data: bytes) -> None:
+        if self._s.pty:
+            self._s.pty.write(data)
+
+    def resize(self, cols: int, rows: int) -> None:
+        if self._s.pty:
+            self._s.pty.resize(cols, rows)
+
+    async def aclose(self) -> None:
+        self._s.subscribers.discard(self._q)
+
+
+class MemoryManager:
+    backend = "memory"
 
     def __init__(self) -> None:
-        self._sessions: dict[str, AgentSession] = {}
+        self._sessions: dict[str, _MemSession] = {}
 
-    def _reap_exited(self) -> None:
+    def _reap(self) -> None:
         now = time.time()
-        stale = [
+        for sid in [
-            sid
+            s.id
-            for sid, s in self._sessions.items()
+            for s in self._sessions.values()
-            if s.exited and s.exited_at and (now - s.exited_at) > _EXITED_TTL
+            if s.exited and s.exited_at and now - s.exited_at > _EXITED_TTL
-        ]
+        ]:
-        for sid in stale:
             self._sessions.pop(sid, None)
 
     async def create(
         self,
         agent: str,
-        command: str,
+        argv: list[str],
-        args: list[str] | None = None,
+        cwd: str,
-        cwd: str | Path | None = None,
+        env: dict[str, str],
-        env: dict[str, str] | None = None,
+        cols: int,
-        cols: int = 80,
+        rows: int,
-        rows: int = 24,
+    ) -> str:
-    ) -> AgentSession:
+        self._reap()
-        self._reap_exited()
+        sid = uuid.uuid4().hex[:12]
-        session = AgentSession(
+        s = _MemSession(sid, agent, shlex.join(argv), cwd)
-            id=uuid.uuid4().hex[:12],
+        s.pty = await PtySession.start(
-            agent=agent,
+            argv[0],
-            command=command,
+            args=argv[1:],
-            cwd=str(cwd) if cwd else "",
+            cwd=cwd or None,
-            created_at=time.time(),
-        )
-        session.pty = await PtySession.start(
-            command,
-            args=args,
-            cwd=cwd,
             env=env,
             cols=cols,
             rows=rows,
-            on_output=session._on_output,
+            on_output=s._on_output,
-            on_exit=session._on_exit,
+            on_exit=s._on_exit,
         )
-        self._sessions[session.id] = session
+        self._sessions[sid] = s
-        return session
+        return sid
 
-    def get(self, session_id: str) -> AgentSession | None:
+    async def get_info(self, sid: str) -> dict | None:
-        return self._sessions.get(session_id)
+        s = self._sessions.get(sid)
+        return s.info() if s else None
 
-    def list(self) -> list[dict]:
+    async def list(self) -> list[dict]:
-        self._reap_exited()
+        self._reap()
         return [
             s.info()
             for s in sorted(
@@ -150,13 +187,25 @@ class SessionManager:
             )
         ]
 
-    async def close(self, session_id: str) -> bool:
+    async def attach(self, sid: str, cols: int, rows: int) -> Attachment | None:
-        session = self._sessions.pop(session_id, None)
+        s = self._sessions.get(sid)
-        if session is None:
+        if s is None or not s.running or s.pty is None:
+            return None
+        # Subscribe + snapshot atomically (no await between → the output callback,
+        # in this same loop, can't interleave), so replay has no gap/dup.
+        q: asyncio.Queue[bytes | None] = asyncio.Queue(maxsize=256)
+        s.subscribers.add(q)
+        snap = bytes(s.scrollback)
+        s.pty.resize(cols, rows)
+        return _MemAttachment(s, q, snap)
+
+    async def close(self, sid: str) -> bool:
+        s = self._sessions.pop(sid, None)
+        if s is None:
             return False
-        if session.pty is not None:
+        if s.pty:
-            await session.pty.close()
+            await s.pty.close()
-        for q in list(session.subscribers):
+        for q in list(s.subscribers):
             q.put_nowait(None)
         return True
 
@@ -165,5 +214,235 @@ class SessionManager:
             await self.close(sid)
 
 
-# Module-level singleton (mirrors mesh_state / stream subscribers).
+# ---------------------------------------------------------------------------
-manager = SessionManager()
+# tmux backend
+# ---------------------------------------------------------------------------
+
+_TMUX_SOCK = CASTLE_HOME / "run" / "agents.sock"
+_PREFIX = "castle-"
+_KEEPALIVE = "__keepalive__"  # hidden session that keeps the isolated server alive
+
+
+class _TmuxAttachment:
+    def __init__(self, pty: PtySession, q: asyncio.Queue[bytes | None]):
+        self._pty = pty
+        self._q = q
+
+    @classmethod
+    async def start(
+        cls, tmux_bin: str, sock: str, name: str, cols: int, rows: int
+    ) -> _TmuxAttachment:
+        q: asyncio.Queue[bytes | None] = asyncio.Queue(maxsize=256)
+        pty = await PtySession.start(
+            tmux_bin,
+            args=["-S", sock, "attach-session", "-t", name],
+            cols=cols,
+            rows=rows,
+            on_output=lambda d: q.put_nowait(d),
+            on_exit=lambda: q.put_nowait(None),
+        )
+        return cls(pty, q)
+
+    async def reader(self) -> AsyncIterator[bytes]:
+        while (chunk := await self._q.get()) is not None:
+            yield chunk
+
+    def write(self, data: bytes) -> None:
+        self._pty.write(data)
+
+    def resize(self, cols: int, rows: int) -> None:
+        self._pty.resize(cols, rows)
+
+    async def aclose(self) -> None:
+        await self._pty.close()  # kills the attach client → detaches; session lives
+
+
+class TmuxManager:
+    backend = "tmux"
+
+    def __init__(self, tmux_bin: str) -> None:
+        self._bin = tmux_bin
+        self._sock = str(_TMUX_SOCK)
+        self._ready = False
+        self._lock = asyncio.Lock()
+
+    async def _raw(self, *args: str) -> tuple[int, str, str]:
+        proc = await asyncio.create_subprocess_exec(
+            self._bin,
+            "-S",
+            self._sock,
+            *args,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+        )
+        out, err = await proc.communicate()
+        return (
+            proc.returncode or 0,
+            out.decode(errors="replace"),
+            err.decode(errors="replace"),
+        )
+
+    async def _server_running(self) -> bool:
+        _, out, err = await self._raw("list-sessions")
+        text = (out + err).lower()
+        return not any(
+            m in text for m in ("no server", "error connecting", "no such file")
+        )
+
+    async def _ensure(self) -> None:
+        if self._ready:
+            return
+        async with self._lock:
+            if self._ready:
+                return
+            _TMUX_SOCK.parent.mkdir(parents=True, exist_ok=True)
+            if not await self._server_running():
+                # Start the server inside its OWN systemd scope so it is NOT in
+                # castle-api's cgroup (which a restart kills). It must be born with
+                # a keepalive session: an empty server races its own empty-exit
+                # check against `exit-empty off` and dies, after which a plain
+                # `new-session` would restart it inside castle-api's cgroup. The
+                # keepalive (a detached `sleep infinity`, name-filtered from
+                # listings) guarantees the isolated server persists.
+                proc = await asyncio.create_subprocess_exec(
+                    "systemd-run",
+                    "--user",
+                    "--scope",
+                    "--quiet",
+                    "--collect",
+                    self._bin,
+                    "-S",
+                    self._sock,
+                    "new-session",
+                    "-d",
+                    "-s",
+                    _KEEPALIVE,
+                    "sleep infinity",
+                    stdout=asyncio.subprocess.DEVNULL,
+                    stderr=asyncio.subprocess.PIPE,
+                )
+                _, err = await proc.communicate()
+                if proc.returncode != 0 or not await self._server_running():
+                    logger.warning(
+                        "agent tmux server not isolated (systemd-run: %s); sessions "
+                        "will still work but won't survive a castle-api restart",
+                        err.decode(errors="replace").strip() or "unknown error",
+                    )
+            # Clean, terminal-like defaults (idempotent).
+            for opt in (
+                ("set", "-g", "status", "off"),  # no tmux status bar
+                ("set", "-g", "window-size", "latest"),  # newest client's size wins
+                ("set", "-g", "escape-time", "0"),
+                ("set", "-g", "destroy-unattached", "off"),
+                ("set", "-s", "exit-empty", "off"),
+            ):
+                await self._raw(*opt)
+            self._ready = True
+
+    async def create(
+        self,
+        agent: str,
+        argv: list[str],
+        cwd: str,
+        env: dict[str, str],
+        cols: int,
+        rows: int,
+    ) -> str:
+        await self._ensure()
+        sid = uuid.uuid4().hex[:12]
+        name = f"{_PREFIX}{sid}"
+        cmd = shlex.join(argv)  # tmux runs this as the pane's shell-command
+        args = [
+            "new-session",
+            "-d",
+            "-s",
+            name,
+            "-x",
+            str(cols),
+            "-y",
+            str(rows),
+        ]
+        if cwd:
+            args += ["-c", cwd]
+        for k, v in env.items():
+            args += ["-e", f"{k}={v}"]
+        args += [cmd]
+        rc, _, err = await self._raw(*args)
+        if rc != 0:
+            raise RuntimeError(f"tmux new-session failed: {err.strip()}")
+        await self._raw("set-option", "-t", name, "@agent", agent)
+        return sid
+
+    async def _list_raw(self) -> list[dict]:
+        fmt = "#{session_name}\t#{session_created}\t#{@agent}\t#{session_attached}\t#{window_width}\t#{window_height}"
+        rc, out, _ = await self._raw("list-sessions", "-F", fmt)
+        if rc != 0:
+            return []
+        rows: list[dict] = []
+        for line in out.splitlines():
+            parts = line.split("\t")
+            if len(parts) < 6 or not parts[0].startswith(_PREFIX):
+                continue
+            name, created, agent, attached, w, h = parts[:6]
+            rows.append(
+                {
+                    "id": name[len(_PREFIX) :],
+                    "agent": agent or name[len(_PREFIX) :],
+                    "command": "",
+                    "cwd": "",
+                    "created_at": float(created) if created.isdigit() else None,
+                    "running": True,
+                    "exited": False,
+                    "exit_code": None,
+                    "cols": int(w) if w.isdigit() else None,
+                    "rows": int(h) if h.isdigit() else None,
+                    "clients": int(attached) if attached.isdigit() else 0,
+                }
+            )
+        return rows
+
+    async def get_info(self, sid: str) -> dict | None:
+        for row in await self._list_raw():
+            if row["id"] == sid:
+                return row
+        return None
+
+    async def list(self) -> list[dict]:
+        rows = await self._list_raw()
+        rows.sort(key=lambda r: r["created_at"] or 0, reverse=True)
+        return rows
+
+    async def attach(self, sid: str, cols: int, rows: int) -> Attachment | None:
+        if await self.get_info(sid) is None:
+            return None
+        return await _TmuxAttachment.start(
+            self._bin, self._sock, f"{_PREFIX}{sid}", cols, rows
+        )
+
+    async def close(self, sid: str) -> bool:
+        rc, _, _ = await self._raw("kill-session", "-t", f"{_PREFIX}{sid}")
+        return rc == 0
+
+    async def close_all(self) -> None:
+        # No-op: tmux sessions are meant to SURVIVE a castle-api shutdown.
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Backend selection
+# ---------------------------------------------------------------------------
+
+
+def _make_manager() -> MemoryManager | TmuxManager:
+    choice = os.environ.get("CASTLE_API_AGENT_BACKEND", "auto").lower()
+    tmux_bin = shutil.which("tmux")
+    if choice == "memory":
+        return MemoryManager()
+    if choice == "tmux" and not tmux_bin:
+        raise RuntimeError("CASTLE_API_AGENT_BACKEND=tmux but tmux is not on PATH")
+    if tmux_bin and choice in ("auto", "tmux"):
+        return TmuxManager(tmux_bin)
+    return MemoryManager()
+
+
+manager: MemoryManager | TmuxManager = _make_manager()

castle-api/src/castle_api/agents.py

@@ -1,8 +1,10 @@
-"""Agent terminal UX — list agents, run/resume them in a pty over WebSocket.
+"""Agent terminal UX — list agents, run/resume them in a terminal over WebSocket.
 
 Castle is assistant-agnostic: it launches a configured command (or a shell) in a
 real tty and streams raw bytes to an xterm.js terminal in the browser. Sessions
-outlive their WebSocket connection, so they can be listed, resumed, and killed.
+outlive their WebSocket connection (and, with the tmux backend, a castle-api
+restart), so they can be listed, resumed, and killed. The session backend lives
+in `agent_sessions.py`; this module is the HTTP/WebSocket surface.
 """
 
 from __future__ import annotations
@@ -24,7 +26,7 @@ from castle_api.agent_registry import (
     resolve_agent,
     resume_argv,
 )
-from castle_api.agent_sessions import AgentSession, manager
+from castle_api.agent_sessions import manager
 from castle_api.config import get_config
 
 logger = logging.getLogger(__name__)
@@ -75,9 +77,9 @@ def get_agents() -> list[dict]:
 
 
 @router.get("/sessions")
-def get_sessions() -> list[dict]:
+async def get_sessions() -> list[dict]:
-    """List live terminal sessions (running or recently exited)."""
+    """List live terminal sessions (running, backend-managed)."""
-    return manager.list()
+    return await manager.list()
 
 
 @router.get("/history")
@@ -97,7 +99,7 @@ async def get_history() -> list[dict]:
 
 @router.delete("/sessions/{session_id}")
 async def delete_session(session_id: str) -> dict:
-    """Kill a session's process group and drop it from the registry."""
+    """Kill a session and drop it from the backend."""
     ok = await manager.close(session_id)
     if not ok:
         raise HTTPException(
@@ -113,121 +115,118 @@ def _login_shell() -> str:
     return os.environ.get("SHELL") or "/bin/bash"
 
 
-async def _new_session(
+async def _build_launch(
     name: str, cont: bool = False, resume_session: str | None = None
-) -> AgentSession:
+) -> tuple[str, list[str], str, dict[str, str]]:
-    """Create a session for a named agent, or the reserved login shell.
+    """Return (agent_label, argv, cwd, env) for a new session.
 
     Everything runs under the user's *interactive login* shell (`-lic`) so it
     inherits exactly the environment a real terminal would — the systemd service
-    env is stripped down and would otherwise miss vars set in shell rc files
+    env is stripped down and would otherwise miss vars from shell rc files (API
-    (API keys, etc.). Agents are `exec`'d so the pty's foreground process is the
+    keys, etc.). Agents are `exec`'d so the terminal's foreground process is the
-    agent itself (clean signals + exit).
+    agent itself. `cont` appends the agent's `resume_args`; `resume_session`
-
+    builds its resume-by-id argv — castle only passes declared flags through.
-    ``cont`` appends the agent's ``resume_args`` (open its own picker / continue).
-    ``resume_session`` builds the agent's resume-by-id argv for a specific past
-    session. Either way castle only passes declared flags through — it never
-    reads the agent's session storage.
     """
     shell = _login_shell()
     if name == TERMINAL_AGENT:
-        return await manager.create(
+        return TERMINAL_AGENT, [shell, "-l"], default_cwd(), {}
-            TERMINAL_AGENT, shell, args=["-l"], cwd=default_cwd()
-        )
     agent = resolve_agent(name)
     if agent is None:
         raise LookupError(f"unknown agent: {name}")
     if not agent.available:
         raise LookupError(f"agent not installed: {name}")
     if resume_session:
-        argv = resume_argv(agent, resume_session)
+        inner = resume_argv(agent, resume_session)
-        if argv is None:
+        if inner is None:
             raise LookupError(f"agent cannot resume by id: {name}")
     else:
-        argv = [agent.resolved_command or agent.command, *agent.args]
+        inner = [agent.resolved_command or agent.command, *agent.args]
         if cont and agent.resume_args:
-            argv += agent.resume_args
+            inner += agent.resume_args
-    launch = shlex.join(argv)
+    argv = [shell, "-lic", f"exec {shlex.join(inner)}"]
-    return await manager.create(
+    return name, argv, agent.cwd, agent.env
-        name,
-        shell,
-        args=["-lic", f"exec {launch}"],
-        cwd=agent.cwd,
-        env=agent.env,  # explicit per-agent overrides; the login shell sets PATH
-    )
 
 
 @router.websocket("/{name}/session")
 async def agent_session(ws: WebSocket, name: str) -> None:
-    """Run (or resume) an agent in a pty; stream it to the browser terminal.
+    """Run (or resume) an agent in a terminal; stream it to the browser.
 
-    Query param `session=<id>` resumes an existing live session (replaying its
+    Query params (mutually exclusive, all optional):
-    scrollback); otherwise a new session is created for `name` (or the reserved
+    - `session=<id>`         resume an existing live session.
-    `terminal` shell). Disconnecting detaches — it does NOT kill the session.
+    - `resume_session=<id>`  start the agent resumed to one of *its own* past
+                             sessions (agent-native, by id).
+    - `mode=continue`        start the agent with its `resume_args`.
+    Disconnecting detaches — it does NOT kill the session.
     """
     if not _origin_allowed(ws.headers.get("origin")):
         await ws.close(code=1008)  # policy violation
         return
 
     resume_id = ws.query_params.get("session")
+    session_id: str
+    is_resume = False
+    agent_label = name
 
-    # Resolve the target session (before accept only for hard rejects).
+    try:
-    session: AgentSession | None = None
+        if resume_id:
-    if resume_id:
+            info = await manager.get_info(resume_id)
-        session = manager.get(resume_id)
+            if not info or not info.get("running"):
-        if session is None or not session.running:
+                await ws.accept()
-            await ws.accept()
+                await ws.send_json({"type": "error", "error": "session not resumable"})
-            await ws.send_json({"type": "error", "error": "session not resumable"})
+                await ws.close()
-            await ws.close()
+                return
-            return
+            session_id = resume_id
-    else:
+            is_resume = True
-        cont = ws.query_params.get("mode") == "continue"
+            agent_label = info.get("agent") or name
-        resume_session = ws.query_params.get("resume_session")
+        else:
-        try:
+            cont = ws.query_params.get("mode") == "continue"
-            session = await _new_session(name, cont=cont, resume_session=resume_session)
+            resume_session = ws.query_params.get("resume_session")
-        except LookupError as e:
+            agent_label, argv, cwd, env = await _build_launch(
-            await ws.accept()
+                name, cont, resume_session
-            await ws.send_json({"type": "error", "error": str(e)})
+            )
-            await ws.close()
+            session_id = await manager.create(
-            return
+                agent_label, argv, cwd, env, cols=80, rows=24
-        except Exception:
+            )
-            logger.exception("failed to launch agent %s", name)
+    except LookupError as e:
-            await ws.accept()
+        await ws.accept()
-            await ws.send_json({"type": "error", "error": "failed to launch"})
+        await ws.send_json({"type": "error", "error": str(e)})
-            await ws.close()
+        await ws.close()
-            return
+        return
+    except Exception:
+        logger.exception("failed to launch agent %s", name)
+        await ws.accept()
+        await ws.send_json({"type": "error", "error": "failed to launch"})
+        await ws.close()
+        return
 
-    assert session is not None  # narrowed by the branches above
     await ws.accept()
-
+    att = await manager.attach(session_id, cols=80, rows=24)
-    # Subscribe + snapshot scrollback atomically (no await between → the output
+    if att is None:
-    # callback, which runs in this same event loop, cannot interleave), so a
+        await ws.send_json({"type": "error", "error": "session not attachable"})
-    # resumed client gets the buffer once with no duplication or gap.
+        await ws.close()
-    q = session.subscribe()
+        return
-    snapshot = bytes(session.scrollback)
 
     await ws.send_json(
         {
             "type": "session",
-            "id": session.id,
+            "id": session_id,
-            "agent": session.agent,
+            "agent": agent_label,
-            "resumed": bool(resume_id),
+            "resumed": is_resume,
         }
     )
-    if snapshot:
-        await ws.send_bytes(snapshot)
 
     async def pump() -> None:
         try:
-            while True:
+            async for chunk in att.reader():
-                chunk = await q.get()
-                if chunk is None:  # child exited
-                    if ws.application_state == WebSocketState.CONNECTED:
-                        await ws.send_json({"type": "exit", "code": session.exit_code})
-                    break
                 if ws.application_state != WebSocketState.CONNECTED:
                     break
                 await ws.send_bytes(chunk)
+            # reader ended without cancellation → the session ended
+            if ws.application_state == WebSocketState.CONNECTED:
+                info = await manager.get_info(session_id)
+                code = info.get("exit_code") if info else None
+                await ws.send_json({"type": "exit", "code": code})
+                await ws.close()
         except Exception:
             pass
 
@@ -239,22 +238,20 @@ async def agent_session(ws: WebSocket, name: str) -> None:
                 break
             data = msg.get("bytes")
             text = msg.get("text")
-            if data is not None and session.pty is not None:
+            if data is not None:
-                session.pty.write(data)
+                att.write(data)
-            elif text is not None and session.pty is not None:
+            elif text is not None:
                 try:
                     ctrl = json.loads(text)
                 except json.JSONDecodeError:
                     continue
                 if ctrl.get("type") == "resize":
-                    session.pty.resize(int(ctrl["cols"]), int(ctrl["rows"]))
+                    att.resize(int(ctrl["cols"]), int(ctrl["rows"]))
                 elif ctrl.get("type") == "input":
-                    session.pty.write(str(ctrl.get("data", "")).encode())
+                    att.write(str(ctrl.get("data", "")).encode())
     except WebSocketDisconnect:
         pass
     finally:
         pump_task.cancel()
-        session.unsubscribe(q)
+        await att.aclose()
-        # Intentionally do NOT close the session here — it lives on for resume.
+        # Intentionally do NOT close the session — it lives on for resume.
-        # Sessions are ended via DELETE /agents/sessions/{id}, when the child
-        # exits, or by idle reaping.