feat(maestro): persist PTY scrollback across a reboot

A resumed session opened on a blank pane: the agent's conversation came
back, the terminal history the operator was reading did not. Each PTY
bridge's replay ring is now mirrored to `<state_dir>/scrollback/<id>.bin`
and handed to the new bridge on resume, and on the recovery pass that
re-adopts a session tmux still holds — that one lost its ring with the
previous daemon too. Restored bytes go out ahead of the live ones, and
are dropped once the ring holds a full budget of new output of its own.

Flushed every 10 seconds by the liveness watcher, and once more when a
bridge is sealed on exit. The timer is the part that matters: a host
reboot runs no exit path at all, so a scrollback written only on a clean
shutdown is missing in exactly the case this exists for. Ten seconds
bounds the loss to the last few lines — this is visual continuity, not
state — and costs at most one file write per active session per tick,
since a session whose bridge has read nothing since the last flush is
skipped.

The file is capped at the existing `pty_replay_bytes` budget rather than
a second one, is deleted on kill and when recovery prunes a sidecar, and
a missing, unreadable or oversized one degrades to an empty pane instead
of failing the resume.

Closes #171

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

crates/maestro-config/src/lib.rs

@@ -129,6 +129,80 @@ pub struct DaemonSection {
     /// `Some(0)` disables history entirely (no log written, no recents read).
     #[serde(default)]
     pub spawn_history_days: Option<u32>,
+    /// What recovery does with sessions that outlived their tmux server —
+    /// after a host reboot, or any restart that finds a sidecar with no pane
+    /// (#167). Default [`ResumeOnBoot::Dormant`].
+    #[serde(default, deserialize_with = "deserialize_resume_on_boot")]
+    pub resume_on_boot: Option<ResumeOnBoot>,
+}
+
+impl DaemonSection {
+    /// The configured `[daemon].resume_on_boot`, or the safe default:
+    /// recovered-but-dead sessions wait for the operator.
+    pub fn resume_on_boot(&self) -> ResumeOnBoot {
+        self.resume_on_boot.unwrap_or_default()
+    }
+}
+
+/// `[daemon].resume_on_boot` — what recovery does with a session whose tmux
+/// pane is gone but whose sidecar is still on disk (#167).
+///
+/// Resuming agents unattended seconds after a reboot is a foot-gun: they run
+/// with whatever flags they were spawned with, with nobody watching. So the
+/// default waits for a human, and `auto` is opted into per host.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub enum ResumeOnBoot {
+    /// Register each one `Dormant` and wait for the operator to resume it.
+    #[default]
+    Dormant,
+    /// Resume each one during recovery, on the same path a client's
+    /// `POST /api/sessions/:id/resume` takes.
+    Auto,
+    /// Prune the sidecars instead: a reboot is a clean slate.
+    Never,
+}
+
+/// Every `[daemon].resume_on_boot` name and its variant, in documentation
+/// order.
+pub const RESUME_ON_BOOT_POLICIES: &[(&str, ResumeOnBoot)] = &[
+    ("dormant", ResumeOnBoot::Dormant),
+    ("auto", ResumeOnBoot::Auto),
+    ("never", ResumeOnBoot::Never),
+];
+
+impl ResumeOnBoot {
+    /// The name this policy is spelled with in the config file and in logs.
+    pub fn as_str(&self) -> &'static str {
+        match self {
+            ResumeOnBoot::Dormant => "dormant",
+            ResumeOnBoot::Auto => "auto",
+            ResumeOnBoot::Never => "never",
+        }
+    }
+}
+
+fn deserialize_resume_on_boot<'de, D>(deserializer: D) -> Result<Option<ResumeOnBoot>, D::Error>
+where
+    D: Deserializer<'de>,
+{
+    let raw = Option::<String>::deserialize(deserializer)?;
+    raw.map(|name| {
+        RESUME_ON_BOOT_POLICIES
+            .iter()
+            .find(|(n, _)| *n == name)
+            .map(|(_, policy)| *policy)
+            .ok_or_else(|| {
+                let known = RESUME_ON_BOOT_POLICIES
+                    .iter()
+                    .map(|(n, _)| format!("{n:?}"))
+                    .collect::<Vec<_>>()
+                    .join(", ");
+                D::Error::custom(format!(
+                    "invalid [daemon].resume_on_boot {name:?}; expected one of {known}"
+                ))
+            })
+    })
+    .transpose()
 }
 
 /// `[tui]` — `maestro tui` settings (NAP-275 epic, phase 2 NAP-277).
@@ -775,4 +849,27 @@ bind_addr = "127.0.0.1:9090"
         );
         assert_eq!(cfg.resolve_daemon_url(None), "http://explicit:1");
     }
+
+    #[test]
+    fn resume_on_boot_defaults_to_dormant_and_parses_every_policy() {
+        assert_eq!(
+            MaestroConfigFile::default().daemon.resume_on_boot(),
+            ResumeOnBoot::Dormant,
+            "an absent policy is the safe one"
+        );
+        for (name, expected) in RESUME_ON_BOOT_POLICIES {
+            let cfg = file_with(&format!("[daemon]\nresume_on_boot = \"{name}\"\n"));
+            assert_eq!(cfg.daemon.resume_on_boot(), *expected, "{name}");
+            assert_eq!(expected.as_str(), *name);
+        }
+    }
+
+    #[test]
+    fn resume_on_boot_rejects_an_unknown_policy() {
+        let err = toml::from_str::<MaestroConfigFile>("[daemon]\nresume_on_boot = \"maybe\"\n")
+            .expect_err("unknown policy must not silently fall back");
+        let msg = err.to_string();
+        assert!(msg.contains("resume_on_boot"), "{msg}");
+        assert!(msg.contains("\"auto\""), "{msg}");
+    }
 }

crates/maestro-daemon/src/boot_id.rs

@@ -2,10 +2,13 @@
 //!
 //! `/proc/sys/kernel/random/boot_id` is fixed for the life of the kernel and
 //! changes on every boot. [`SessionManager::recover`] compares it against the
-//! value it stored on the previous run: a difference means every session from
-//! before is provably dead, so the tmux probe can be skipped and those
-//! sessions registered dormant directly instead of relying on tmux reporting
-//! an empty list.
+//! value it stored on the previous run and logs which of the two it was.
+//!
+//! It does not decide liveness (#174). A boot id changes without the tmux
+//! server dying whenever the state directory is restored, a VM or container
+//! snapshot is rolled back, or the `boot_id` file is truncated — so recovery
+//! always probes `tmux list-sessions` and the boot id only says whether a
+//! server that did not answer is coming back.
 
 use std::path::Path;
 

crates/maestro-daemon/src/cli.rs

@@ -84,10 +84,16 @@ async fn run(config_override: Option<PathBuf>) -> Result<(), Box<dyn Error>> {
         config.spawn_history_path(),
         config.spawn_history_days,
     ))
+    // #171: mirror PTY scrollback so a resumed session opens on what it was
+    // showing, instead of a blank pane.
+    .with_scrollback(config.scrollback_dir())
     .with_boot_id(
         config.boot_id_path(),
         std::sync::Arc::new(crate::boot_id::ProcBootId),
-    );
+    )
+    // #167: recovery needs the presets to resolve a resumed session's command,
+    // so `auto` runs the same resume a client's POST would.
+    .with_resume_policy(config.resume_on_boot, config.spawn_presets.clone());
     // Spawned sessions learn where to report back (#103). Derived from the
     // configured bind address; set before any spawn can happen.
     manager.set_daemon_url(format!("http://{}", config.bind_addr));

crates/maestro-daemon/src/config.rs

@@ -97,6 +97,10 @@ pub struct DaemonConfig {
     /// never binds it itself; it serves it so the web client and the TUI read
     /// one knob.
     pub leader: String,
+    /// What recovery does with sessions whose tmux pane is gone but whose
+    /// sidecar survives (#167): `[daemon].resume_on_boot`, default
+    /// [`maestro_config::ResumeOnBoot::Dormant`].
+    pub resume_on_boot: maestro_config::ResumeOnBoot,
 }
 
 impl DaemonConfig {
@@ -113,6 +117,12 @@ impl DaemonConfig {
         self.state_dir.join("spawn-history.jsonl")
     }
 
+    /// Where PTY scrollback is mirrored so a resumed session opens on what
+    /// it printed before (#171): `<state_dir>/scrollback`.
+    pub fn scrollback_dir(&self) -> PathBuf {
+        self.state_dir.join("scrollback")
+    }
+
     /// Where recovery records the kernel boot id it saw last (#164):
     /// `<state_dir>/boot_id`.
     pub fn boot_id_path(&self) -> PathBuf {
@@ -149,6 +159,8 @@ pub fn load_config(path: &Path, env: EnvOverrides) -> Result<DaemonConfig, Confi
     // validation (#22/#26).
     let spawn_presets = file.spawn.resolve()?;
 
+    let resume_on_boot = daemon.resume_on_boot();
+
     let hostname = daemon.hostname.unwrap_or_else(system_hostname);
     let home_dir = dirs::home_dir()
         .map(|p| p.to_string_lossy().into_owned())
@@ -209,6 +221,7 @@ pub fn load_config(path: &Path, env: EnvOverrides) -> Result<DaemonConfig, Confi
         project_scan_depth,
         hooks_auto_install: file.hooks.auto_install(),
         leader: file.tui.leader().to_string(),
+        resume_on_boot,
     })
 }
 
@@ -289,6 +302,34 @@ auto_install = false
         assert!(!cfg.hooks_auto_install);
     }
 
+    #[test]
+    fn resume_on_boot_defaults_to_dormant_and_reads_the_file() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("config.toml");
+        std::fs::write(&path, "").unwrap();
+        assert_eq!(
+            load_config(&path, EnvOverrides::default())
+                .unwrap()
+                .resume_on_boot,
+            maestro_config::ResumeOnBoot::Dormant
+        );
+
+        std::fs::write(
+            &path,
+            r#"
+[daemon]
+resume_on_boot = "auto"
+"#,
+        )
+        .unwrap();
+        assert_eq!(
+            load_config(&path, EnvOverrides::default())
+                .unwrap()
+                .resume_on_boot,
+            maestro_config::ResumeOnBoot::Auto
+        );
+    }
+
     #[test]
     fn empty_file_defaults_everything() {
         let dir = tempfile::tempdir().unwrap();

crates/maestro-daemon/src/lib.rs

@@ -26,6 +26,7 @@ pub mod projects;
 pub mod pty;
 pub mod release;
 pub mod resume;
+pub mod scrollback;
 pub mod session;
 pub mod sidecar;
 pub mod spawn_history;

crates/maestro-daemon/src/pty.rs

@@ -21,6 +21,11 @@
 //!   one-shot prefix before joining the live broadcast, so attaching to an
 //!   idle session immediately shows the last screenful instead of a blank
 //!   terminal.
+//! - #171: a bridge can be started with scrollback restored from disk — what
+//!   the session printed before the daemon (or the host) went away. Those
+//!   bytes are handed to subscribers ahead of the live ones until the ring
+//!   has a full budget of new output of its own, at which point the restored
+//!   history is older than anything the budget covers and is dropped.
 
 use std::collections::VecDeque;
 use std::ffi::CString;
@@ -58,6 +63,10 @@ const READ_BUF_BYTES: usize = 4096;
 #[derive(Debug)]
 pub struct PtyBridge {
     fifo_path: PathBuf,
+    /// Scrollback from a previous life of this session (#171), replayed
+    /// ahead of the ring. Empty for a fresh spawn. Never appended to, so it
+    /// can never duplicate a byte the ring also holds.
+    restored: Vec<u8>,
     /// `None` once [`Self::seal`] has run: live subscribers see the channel
     /// close, but the replay buffer stays readable.
     sender: Option<broadcast::Sender<Vec<u8>>>,
@@ -79,6 +88,21 @@ impl PtyBridge {
     /// Variant that lets the caller pick a custom replay-buffer cap (bytes).
     /// Pass `0` to disable replay (subscribers see only live bytes).
     pub async fn start_with_replay(fifo_path: PathBuf, replay_bytes: usize) -> io::Result<Self> {
+        Self::start_restoring(fifo_path, replay_bytes, Vec::new()).await
+    }
+
+    /// As [`Self::start_with_replay`], seeded with `restored` scrollback read
+    /// back from disk (#171). The seed is truncated to the same
+    /// `replay_bytes` budget the ring uses — there is one budget here, not
+    /// two — and is dropped entirely when replay is disabled.
+    pub async fn start_restoring(
+        fifo_path: PathBuf,
+        replay_bytes: usize,
+        mut restored: Vec<u8>,
+    ) -> io::Result<Self> {
+        if restored.len() > replay_bytes {
+            restored.drain(..restored.len() - replay_bytes);
+        }
         if let Some(parent) = fifo_path.parent() {
             tokio::fs::create_dir_all(parent).await?;
         }
@@ -102,6 +126,7 @@ impl PtyBridge {
         });
         Ok(Self {
             fifo_path,
+            restored,
             sender: Some(sender),
             replay,
             last_activity,
@@ -155,7 +180,8 @@ impl PtyBridge {
         F: FnOnce() -> Option<Vec<u8>>,
     {
         let guard = self.replay.lock_for_snapshot();
-        let snapshot = capture().unwrap_or_else(|| guard.snapshot());
+        let live = capture().unwrap_or_else(|| guard.snapshot());
+        let snapshot = guard.with_restored_prefix(&self.restored, live);
         let rx = match &self.sender {
             Some(tx) => tx.subscribe(),
             // Sealed: hand back the scrollback plus a receiver that is already
@@ -171,6 +197,25 @@ impl PtyBridge {
         (snapshot, rx)
     }
 
+    /// Everything worth persisting for a later life of this session (#171):
+    /// the restored scrollback and the ring, trimmed together to the replay
+    /// budget so the file on disk is never bigger than the ring it mirrors.
+    pub fn scrollback(&self) -> Vec<u8> {
+        let guard = self.replay.lock_for_snapshot();
+        guard.with_restored_prefix(&self.restored, guard.snapshot())
+    }
+
+    /// Total bytes this bridge has appended to its ring. The scrollback flush
+    /// pass compares it against what it last wrote, so an idle session costs
+    /// no disk writes at all.
+    pub fn bytes_seen(&self) -> u64 {
+        self.replay
+            .inner
+            .lock()
+            .expect("ReplayBuffer poisoned")
+            .seen
+    }
+
     /// Stop reading and close the live stream, keeping the replay buffer.
     ///
     /// Attached clients see the channel close — without that a `maestro attach`
@@ -201,6 +246,8 @@ struct ReplayBuffer {
 struct ReplayInner {
     buf: VecDeque<u8>,
     cap: usize,
+    /// Bytes ever appended, not bytes held — the ring evicts, this does not.
+    seen: u64,
 }
 
 impl ReplayBuffer {
@@ -209,6 +256,7 @@ impl ReplayBuffer {
             inner: Arc::new(Mutex::new(ReplayInner {
                 buf: VecDeque::with_capacity(cap.min(READ_BUF_BYTES * 2)),
                 cap,
+                seen: 0,
             })),
         }
     }
@@ -225,6 +273,7 @@ impl ReplayBuffer {
         // If the incoming chunk alone exceeds the cap, take only its tail —
         // anything in the buffer would be evicted before this finishes anyway.
         let take_from = bytes.len().saturating_sub(s.cap);
+        s.seen += bytes.len() as u64;
         let bytes = &bytes[take_from..];
         s.buf.extend(bytes);
         while s.buf.len() > s.cap {
@@ -247,6 +296,24 @@ struct ReplayGuard<'a> {
 }
 
 impl ReplayGuard<'_> {
+    /// `live` with the restored scrollback (#171) in front of it, while the
+    /// ring still holds less than a full budget of its own output. Once it
+    /// holds a full budget, everything restored is older than what the
+    /// budget covers, so it is dropped rather than kept forever.
+    fn with_restored_prefix(&self, restored: &[u8], live: Vec<u8>) -> Vec<u8> {
+        let cap = self.inner.cap;
+        let held = self.inner.buf.len();
+        if restored.is_empty() || cap == 0 || held >= cap {
+            return live;
+        }
+        let keep = cap - held;
+        let start = restored.len().saturating_sub(keep);
+        let mut out = Vec::with_capacity((restored.len() - start) + live.len());
+        out.extend_from_slice(&restored[start..]);
+        out.extend_from_slice(&live);
+        out
+    }
+
     fn snapshot(&self) -> Vec<u8> {
         let (a, b) = self.inner.buf.as_slices();
         let mut out = Vec::with_capacity(a.len() + b.len());
@@ -292,6 +359,7 @@ async fn run_reader(
                 // never both, never neither.
                 let mut s = replay.inner.lock().expect("ReplayBuffer poisoned");
                 if s.cap > 0 {
+                    s.seen += n as u64;
                     let take_from = n.saturating_sub(s.cap);
                     s.buf.extend(&buf[take_from..n]);
                     while s.buf.len() > s.cap {
@@ -345,6 +413,80 @@ fn shell_quote(s: &str) -> String {
 mod tests {
     use super::*;
 
+    async fn bridge(cap: usize, restored: &[u8]) -> (tempfile::TempDir, PtyBridge) {
+        let dir = tempfile::TempDir::new().expect("tempdir");
+        let bridge =
+            PtyBridge::start_restoring(dir.path().join("pty.fifo"), cap, restored.to_vec())
+                .await
+                .expect("start bridge");
+        (dir, bridge)
+    }
+
+    /// Feed the bridge through its own FIFO and wait for the reader to have
+    /// taken the bytes, so the ring state is deterministic.
+    async fn feed(bridge: &PtyBridge, bytes: &[u8]) {
+        use std::io::Write;
+        let before = bridge.bytes_seen();
+        let mut f = std::fs::OpenOptions::new()
+            .write(true)
+            .open(bridge.fifo_path())
+            .expect("open fifo for write");
+        f.write_all(bytes).expect("write fifo");
+        drop(f);
+        for _ in 0..400 {
+            if bridge.bytes_seen() >= before + bytes.len() as u64 {
+                return;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(5)).await;
+        }
+        panic!("bridge never read the bytes written to its fifo");
+    }
+
+    #[tokio::test]
+    async fn restored_scrollback_is_replayed_ahead_of_live_bytes() {
+        let (_dir, bridge) = bridge(64, b"before the reboot\n").await;
+        feed(&bridge, b"after\n").await;
+        let (replay, _rx) = bridge.subscribe_with_replay();
+        assert_eq!(replay, b"before the reboot\nafter\n");
+    }
+
+    #[tokio::test]
+    async fn restored_scrollback_is_trimmed_to_the_replay_budget() {
+        let (_dir, bridge) = bridge(4, b"abcdefghij").await;
+        let (replay, _rx) = bridge.subscribe_with_replay();
+        assert_eq!(replay, b"ghij");
+    }
+
+    #[tokio::test]
+    async fn restored_scrollback_yields_once_the_ring_holds_a_full_budget() {
+        let (_dir, bridge) = bridge(4, b"old!").await;
+        feed(&bridge, b"ab").await;
+        let (replay, _rx) = bridge.subscribe_with_replay();
+        assert_eq!(replay, b"d!ab", "prefix plus ring stays inside the budget");
+        feed(&bridge, b"cd").await;
+        let (replay, _rx) = bridge.subscribe_with_replay();
+        assert_eq!(
+            replay, b"abcd",
+            "restored history is now older than the budget"
+        );
+    }
+
+    #[tokio::test]
+    async fn scrollback_carries_the_restored_prefix_and_the_ring() {
+        let (_dir, bridge) = bridge(8, b"old").await;
+        feed(&bridge, b"new").await;
+        assert_eq!(bridge.scrollback(), b"oldnew");
+        assert_eq!(bridge.bytes_seen(), 3);
+    }
+
+    #[tokio::test]
+    async fn no_replay_budget_means_no_restored_scrollback() {
+        let (_dir, bridge) = bridge(0, b"old").await;
+        let (replay, _rx) = bridge.subscribe_with_replay();
+        assert!(replay.is_empty());
+        assert!(bridge.scrollback().is_empty());
+    }
+
     #[test]
     fn replay_buffer_keeps_last_cap_bytes() {
         let r = ReplayBuffer::new(4);

crates/maestro-daemon/src/scrollback.rs

@@ -0,0 +1,148 @@
+//! Persisted PTY scrollback (#171).
+//!
+//! The PTY replay ring lives in memory, so everything a session printed is
+//! gone the moment the daemon process is. A resumed session then opens on a
+//! blank pane: the agent's conversation is back, but the terminal history the
+//! operator was reading is not.
+//!
+//! This store mirrors each bridge's ring to `<state_dir>/scrollback/<id>.bin`,
+//! and hands it back when a session gets a new bridge — on resume, and on the
+//! recovery pass that re-adopts a session tmux still holds. The file is
+//! capped at the same `replay_bytes` budget the in-memory ring uses, so disk
+//! cost is bounded by the same number the operator already configures.
+//!
+//! Nothing here is allowed to fail a resume. The payload is raw PTY bytes
+//! with no framing to be corrupt, and a file that is missing, unreadable or
+//! oversized degrades to an empty (or truncated) pane rather than an error.
+
+use std::io;
+use std::path::{Path, PathBuf};
+
+use uuid::Uuid;
+
+/// Filesystem-backed scrollback store, rooted at a directory.
+#[derive(Debug, Clone)]
+pub struct ScrollbackStore {
+    dir: PathBuf,
+}
+
+impl ScrollbackStore {
+    pub fn new(dir: impl Into<PathBuf>) -> Self {
+        Self { dir: dir.into() }
+    }
+
+    pub fn dir(&self) -> &Path {
+        &self.dir
+    }
+
+    pub fn path_for(&self, id: Uuid) -> PathBuf {
+        self.dir.join(format!("{id}.bin"))
+    }
+
+    /// Atomically write `bytes` to `<dir>/<id>.bin`, creating the directory if
+    /// needed. Written through a temp file and renamed, so a reboot in the
+    /// middle of a flush leaves the previous scrollback intact rather than a
+    /// half-file.
+    pub fn write(&self, id: Uuid, bytes: &[u8]) -> io::Result<()> {
+        std::fs::create_dir_all(&self.dir)?;
+        let tmp = self.dir.join(format!("{id}.bin.tmp"));
+        std::fs::write(&tmp, bytes)?;
+        std::fs::rename(&tmp, self.path_for(id))
+    }
+
+    /// The stored scrollback for `id`, capped to the last `budget` bytes.
+    ///
+    /// Never fails: a missing file, an unreadable one, or a `budget` of zero
+    /// all read as "no scrollback", which is exactly the blank pane the
+    /// operator had before this existed.
+    pub fn read(&self, id: Uuid, budget: usize) -> Vec<u8> {
+        if budget == 0 {
+            return Vec::new();
+        }
+        let path = self.path_for(id);
+        let mut bytes = match std::fs::read(&path) {
+            Ok(b) => b,
+            Err(e) if e.kind() == io::ErrorKind::NotFound => return Vec::new(),
+            Err(e) => {
+                tracing::warn!(error = ?e, %id, path = %path.display(), "scrollback: unreadable, resuming with a blank pane");
+                return Vec::new();
+            }
+        };
+        // A file written under a larger budget than this daemon now runs
+        // with keeps its tail, the same end of the stream the ring would.
+        if bytes.len() > budget {
+            bytes.drain(..bytes.len() - budget);
+        }
+        bytes
+    }
+
+    /// Delete the scrollback for `id`. No-op if it is not there.
+    pub fn delete(&self, id: Uuid) -> io::Result<()> {
+        match std::fs::remove_file(self.path_for(id)) {
+            Ok(()) => Ok(()),
+            Err(e) if e.kind() == io::ErrorKind::NotFound => Ok(()),
+            Err(e) => Err(e),
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn store() -> (tempfile::TempDir, ScrollbackStore) {
+        let dir = tempfile::TempDir::new().expect("tempdir");
+        let store = ScrollbackStore::new(dir.path().join("scrollback"));
+        (dir, store)
+    }
+
+    #[test]
+    fn round_trips_bytes() {
+        let (_dir, store) = store();
+        let id = Uuid::new_v4();
+        store.write(id, b"hello world").expect("write");
+        assert_eq!(store.read(id, 64), b"hello world");
+    }
+
+    #[test]
+    fn a_missing_file_reads_as_empty() {
+        let (_dir, store) = store();
+        assert!(store.read(Uuid::new_v4(), 64).is_empty());
+    }
+
+    #[test]
+    fn a_file_over_budget_keeps_its_tail() {
+        let (_dir, store) = store();
+        let id = Uuid::new_v4();
+        store.write(id, b"abcdefghij").expect("write");
+        assert_eq!(store.read(id, 4), b"ghij");
+    }
+
+    #[test]
+    fn a_zero_budget_reads_as_empty() {
+        let (_dir, store) = store();
+        let id = Uuid::new_v4();
+        store.write(id, b"abcdefghij").expect("write");
+        assert!(store.read(id, 0).is_empty());
+    }
+
+    #[test]
+    fn delete_removes_the_file_and_tolerates_a_missing_one() {
+        let (_dir, store) = store();
+        let id = Uuid::new_v4();
+        store.write(id, b"gone soon").expect("write");
+        store.delete(id).expect("delete");
+        assert!(!store.path_for(id).exists());
+        store.delete(id).expect("second delete is a no-op");
+    }
+
+    #[test]
+    fn a_second_write_replaces_the_first() {
+        let (_dir, store) = store();
+        let id = Uuid::new_v4();
+        store.write(id, b"first").expect("write");
+        store.write(id, b"second").expect("rewrite");
+        assert_eq!(store.read(id, 64), b"second");
+        assert!(!store.dir().join(format!("{id}.bin.tmp")).exists());
+    }
+}

crates/maestro-daemon/src/session.rs

@@ -12,6 +12,7 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use chrono::Utc;
+use maestro_config::ResumeOnBoot;
 use maestro_protocol::{
     ActivityState, CatalogEvent, CreateSessionRequest, ErrorInfo, NameSource, Session,
     SessionStatus, WorktreeSpec, NAME_KEY, NAME_SOURCE_KEY,
@@ -22,6 +23,7 @@ use uuid::Uuid;
 use crate::boot_id::BootIdSource;
 use crate::panes::Panes;
 use crate::pty::{pipe_pane_command, PtyBridge};
+use crate::scrollback::ScrollbackStore;
 use crate::sidecar::{SidecarError, SidecarStore};
 use crate::spawn_history::{RecentDir, SpawnHistory, SpawnHistoryRecord};
 use crate::tmux::TmuxError;
@@ -43,6 +45,16 @@ pub const DEFAULT_LIVENESS_INTERVAL: Duration = Duration::from_secs(2);
 /// attempt, which is all #83 needs.
 pub const DEFAULT_BLOCKED_WORK_INTERVAL: Duration = Duration::from_secs(15);
 
+/// How often the liveness watcher mirrors PTY scrollback to disk (#171).
+///
+/// A host reboot gives no warning and runs no exit path, so this interval is
+/// the whole guarantee: whatever a session printed in the last few seconds
+/// before the power went is what a resumed pane is missing. Ten seconds keeps
+/// that loss cosmetic — scrollback is visual continuity, not state — while
+/// costing at most one file write per *active* session per tick, since a
+/// session whose bridge read nothing is skipped entirely.
+pub const DEFAULT_SCROLLBACK_INTERVAL: Duration = Duration::from_secs(10);
+
 /// Gap between the text of a `send` and the Enter that submits it.
 ///
 /// Agents that run a full-screen prompt (Claude Code, opencode) coalesce a
@@ -86,11 +98,20 @@ pub struct SessionRecord {
     /// (`POST /.../focus`), `None` while unfocused. Daemon-local; never rides
     /// the wire.
     pub focused_since: Option<std::time::Instant>,
+    /// Set for as long as a [`SessionManager::resume`] is between its two
+    /// takes of the map lock (#175). The spawn happens with the lock released,
+    /// so the `Running` check alone lets a second request through; this is the
+    /// state the second one is refused on. Daemon-local; never rides the wire.
+    pub resuming: bool,
     /// Sticky acknowledgement applied by [`crate::activity::apply_focus_suppression`]:
     /// once the pane has been focused long enough to clear an `AwaitingInput`
     /// glow, this holds through focus loss until the session does new work.
     /// Daemon-local; never rides the wire.
     pub awaiting_suppressed: bool,
+    /// [`PtyBridge::bytes_seen`] as of this session's last scrollback flush
+    /// (#171). The flush pass skips a session whose bridge has read nothing
+    /// since, so an idle session costs no disk writes.
+    pub flushed_bytes: u64,
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -195,6 +216,15 @@ struct Inner {
     /// reboot detection is unwired, which behaves exactly like a boot id
     /// that could never be read: always fall back to probing tmux.
     boot_id: Option<BootIdConfig>,
+    /// What recovery does with a session whose pane is gone (#167), and the
+    /// presets an `auto` resume resolves its command from. Defaults to
+    /// [`ResumeOnBoot::Dormant`] with no presets: register and wait.
+    resume_policy: ResumeOnBoot,
+    resume_presets: Vec<maestro_protocol::SpawnPreset>,
+    /// Where PTY scrollback is mirrored so it survives the daemon, and the
+    /// host (#171). `None` — the default — means scrollback is not persisted
+    /// at all: a resumed session opens on a blank pane, exactly as before.
+    scrollback: Option<ScrollbackStore>,
 }
 
 /// Where recovery persists the boot id it last saw, and how it reads the
@@ -253,6 +283,9 @@ impl SessionManager {
                 viewports: Mutex::new(Viewports::new()),
                 next_viewport: AtomicU64::new(0),
                 boot_id: None,
+                resume_policy: ResumeOnBoot::default(),
+                resume_presets: Vec::new(),
+                scrollback: None,
             }),
         }
     }
@@ -290,6 +323,21 @@ impl SessionManager {
     /// # Panics
     ///
     /// If the handle has already been cloned.
+    /// Persist PTY scrollback under `dir` (#171), so a resumed session comes
+    /// back showing what it printed before the daemon — or the host — went
+    /// away. Builder, called once right after construction like the other
+    /// `with_*` methods. Without it nothing is written, read or deleted.
+    ///
+    /// # Panics
+    ///
+    /// If the handle has already been cloned; see [`Self::with_spawn_history`].
+    pub fn with_scrollback(mut self, dir: PathBuf) -> Self {
+        let inner = Arc::get_mut(&mut self.inner)
+            .expect("with_scrollback must be called before the SessionManager is cloned");
+        inner.scrollback = Some(ScrollbackStore::new(dir));
+        self
+    }
+
     pub fn with_boot_id(mut self, path: PathBuf, source: Arc<dyn BootIdSource>) -> Self {
         let inner = Arc::get_mut(&mut self.inner)
             .expect("with_boot_id must be called before the SessionManager is cloned");
@@ -297,6 +345,26 @@ impl SessionManager {
         self
     }
 
+    /// Wire the `[daemon].resume_on_boot` policy (#167) and the spawn presets
+    /// an `auto` resume resolves each session's command from. Builder, called
+    /// once right after construction like the other `with_*` methods. Without
+    /// it, recovery behaves as [`ResumeOnBoot::Dormant`].
+    ///
+    /// # Panics
+    ///
+    /// If the handle has already been cloned.
+    pub fn with_resume_policy(
+        mut self,
+        policy: ResumeOnBoot,
+        presets: Vec<maestro_protocol::SpawnPreset>,
+    ) -> Self {
+        let inner = Arc::get_mut(&mut self.inner)
+            .expect("with_resume_policy must be called before the SessionManager is cloned");
+        inner.resume_policy = policy;
+        inner.resume_presets = presets;
+        self
+    }
+
     /// Set the daemon's own HTTP base URL (#103), exported to every session
     /// spawned afterwards as `MAESTRO_URL`. Unlike the `with_*` builders this
     /// is a plain setter: the daemon runner learns its bound address after
@@ -567,6 +635,8 @@ impl SessionManager {
                     naming: false,
                     focused_since: None,
                     awaiting_suppressed: false,
+                    resuming: false,
+                    flushed_bytes: 0,
                 },
             );
         }
@@ -618,15 +688,23 @@ impl SessionManager {
         id: Uuid,
         presets: &[maestro_protocol::SpawnPreset],
     ) -> Result<Session, SessionError> {
+        // The spawn below runs with the map lock released, so the `Running`
+        // check alone is not enough to keep two resumes apart (#175): both
+        // would pass it, both would open a `PtyBridge` on the same FIFO, and
+        // the loser would report a tmux failure instead of the conflict it
+        // is. The record is marked in-flight under this first lock, and every
+        // path out of here clears it again.
         let mut session = {
-            let sessions = self.inner.sessions.lock().await;
-            let rec = sessions.get(&id).ok_or(SessionError::NotFound(id))?;
-            if rec.session.status == SessionStatus::Running {
+            let mut sessions = self.inner.sessions.lock().await;
+            let rec = sessions.get_mut(&id).ok_or(SessionError::NotFound(id))?;
+            if rec.session.status == SessionStatus::Running || rec.resuming {
                 return Err(SessionError::AlreadyRunning(id));
             }
+            rec.resuming = true;
             rec.session.clone()
         };
         if !session.cwd.exists() {
+            self.clear_resuming(id).await;
             return Err(SessionError::CwdMissing(session.cwd.clone()));
         }
         let command = crate::resume::resume_command(&session, presets);
@@ -634,8 +712,14 @@ impl SessionManager {
         // Same fifo path as the original spawn: it is derived from the id, and
         // `mkfifo` tolerates the stale node the previous run left behind.
         let fifo_path = self.inner.fifo_dir.join(format!("{id}.fifo"));
-        let bridge =
-            PtyBridge::start_with_replay(fifo_path.clone(), self.inner.replay_bytes).await?;
+        // #171: the pane is new, but the scrollback the operator was reading
+        // before the reboot is on disk. It goes in ahead of the live bytes.
+        let bridge = PtyBridge::start_restoring(
+            fifo_path.clone(),
+            self.inner.replay_bytes,
+            self.restored_scrollback(id),
+        )
+        .await?;
         let mut session_env = sanitize_env(&session.env)?;
         session_env.insert("MAESTRO_SESSION_ID".into(), id.to_string());
         if let Some(url) = self
@@ -658,6 +742,7 @@ impl SessionManager {
         if let Err(e) = spawned {
             // `bridge` drops here, closing the FIFO reader. Nothing else was
             // touched, so the session stays exactly as dormant as it was.
+            self.clear_resuming(id).await;
             return Err(SessionError::Tmux(e));
         }
 
@@ -677,6 +762,7 @@ impl SessionManager {
             // sidecar does not describe.
             let tmux = self.inner.tmux.clone();
             let _ = tmux_blocking(move || tmux.kill_session(id)).await;
+            self.clear_resuming(id).await;
             return Err(SessionError::Sidecar(e));
         }
 
@@ -684,9 +770,17 @@ impl SessionManager {
             let mut sessions = self.inner.sessions.lock().await;
             let Some(rec) = sessions.get_mut(&id) else {
                 // Killed while we were spawning. Undo the pane instead of
-                // leaving one behind with no record of it.
+                // leaving one behind with no record of it — and delete the
+                // sidecar the write above just put back (#175). The kill
+                // deleted it before that write landed, so leaving it would
+                // have the next recovery register a dormant ghost for a
+                // session that is gone and whose worktree is already removed.
+                drop(sessions);
                 let tmux = self.inner.tmux.clone();
                 let _ = tmux_blocking(move || tmux.kill_session(id)).await;
+                if let Err(e) = self.inner.sidecar.delete(id) {
+                    tracing::warn!(error = ?e, %id, "resume: failed to delete the sidecar of a session killed mid-resume");
+                }
                 return Err(SessionError::NotFound(id));
             };
             rec.session = session.clone();
@@ -696,6 +790,7 @@ impl SessionManager {
             rec.settled = false;
             rec.focused_since = None;
             rec.awaiting_suppressed = false;
+            rec.resuming = false;
         }
         let _ = self.inner.events.send(CatalogEvent::SessionUpdated {
             session: session.clone(),
@@ -703,6 +798,53 @@ impl SessionManager {
         Ok(session)
     }
 
+    /// Delete the scrollback of a sidecar recovery threw away (#171): the
+    /// session will never be listed again, so nothing would ever read it.
+    fn prune_scrollback(&self, id: Uuid) {
+        if let Some(store) = &self.inner.scrollback {
+            if let Err(e) = store.delete(id) {
+                tracing::warn!(error = ?e, %id, "recover: failed to prune scrollback");
+            }
+        }
+    }
+
+    /// Scrollback persisted for `id` by an earlier life of this session
+    /// (#171), capped to the replay budget. Empty when scrollback is not
+    /// wired, and empty — never an error — when the file is missing or
+    /// unreadable: a blank pane is the honest fallback, and it is what the
+    /// operator had before any of this existed.
+    fn restored_scrollback(&self, id: Uuid) -> Vec<u8> {
+        self.inner
+            .scrollback
+            .as_ref()
+            .map(|store| store.read(id, self.inner.replay_bytes))
+            .unwrap_or_default()
+    }
+
+    /// Mirror every live bridge's scrollback to disk (#171).
+    ///
+    /// Called on a timer by the liveness watcher rather than only on a clean
+    /// exit, because the case this exists for — a host reboot — runs no exit
+    /// path at all: anything flushed only on shutdown is missing exactly when
+    /// it is needed. Sessions whose bridge has read nothing since the last
+    /// flush are skipped, so an idle host writes nothing.
+    ///
+    /// Best-effort throughout: a failed write is warned and the session keeps
+    /// running.
+    pub async fn flush_scrollback(&self) {
+        run_scrollback_pass(&self.inner).await;
+    }
+
+    /// Drop the in-flight marker [`Self::resume`] set under its first lock
+    /// (#175), so a refused or failed resume can be retried. A no-op when the
+    /// record is already gone: a kill that landed mid-resume took the marker
+    /// with it.
+    async fn clear_resuming(&self, id: Uuid) {
+        if let Some(rec) = self.inner.sessions.lock().await.get_mut(&id) {
+            rec.resuming = false;
+        }
+    }
+
     /// Best-effort removal of a worktree created earlier in `create` when a
     /// later step fails. No-op when the request had no worktree.
     fn rollback_worktree(&self, base: &Option<PathBuf>, path: &Option<PathBuf>) {
@@ -793,6 +935,10 @@ impl SessionManager {
             // Start "due" so the first live tick probes: a recovered daemon
             // should not wait a full interval to show what blocks.
             let mut since_probe = blocked_work_interval;
+            // Scrollback starts *not* due: a daemon that just recovered has
+            // nothing new to mirror, and the sessions it adopted were primed
+            // from the files this would overwrite (#171).
+            let mut since_flush = Duration::ZERO;
             loop {
                 ticker.tick().await;
                 let Some(inner) = weak.upgrade() else {
@@ -807,6 +953,11 @@ impl SessionManager {
                     since_probe = Duration::ZERO;
                     run_blocked_work_pass(&inner).await;
                 }
+                since_flush += interval;
+                if since_flush >= DEFAULT_SCROLLBACK_INTERVAL {
+                    since_flush = Duration::ZERO;
+                    run_scrollback_pass(&inner).await;
+                }
             }
         });
     }
@@ -819,20 +970,20 @@ impl SessionManager {
     /// that no longer exists are still adopted — the spec does not require
     /// us to repair them.
     ///
-    /// #164: when the kernel's boot id has changed since the last run, the
-    /// tmux probe below is skipped entirely — a reboot kills the tmux server,
-    /// so every previous session is already dead and there is nothing tmux
-    /// could truthfully adopt. [`Self::register_dormant`] then registers all
-    /// of them directly.
+    /// #164/#174: the kernel's boot id says whether the host rebooted since
+    /// the last run, but it never decides which sessions are alive —
+    /// `tmux list-sessions` does, always, which is what the rest of the
+    /// daemon assumes. A boot id can change without the tmux server dying
+    /// (a restored state directory, a VM or container snapshot restore, a
+    /// truncated `boot_id` file); skipping the probe then registered every
+    /// live session `Dormant`, and resume could never bring one back because
+    /// the tmux name was still taken. The boot id is now only a signal that a
+    /// tmux server which did not answer is not coming back, and it names
+    /// itself in recovery's log.
     pub async fn recover(&self) -> Result<usize, SessionError> {
         tokio::fs::create_dir_all(&self.inner.fifo_dir).await?;
         let rebooted = self.check_boot_id().await;
-        let live = if rebooted {
-            // #164: every session from the previous boot is provably dead —
-            // no tmux server survives a reboot — so there is nothing to gain
-            // from asking it and no session it could honestly report back.
-            Vec::new()
-        } else {
+        let live = {
             let tmux = self.inner.tmux.clone();
             tmux_blocking(move || tmux.list_sessions()).await?
         };
@@ -866,8 +1017,15 @@ impl SessionManager {
                 },
             };
             let fifo_path = self.inner.fifo_dir.join(format!("{id}.fifo"));
-            let bridge =
-                PtyBridge::start_with_replay(fifo_path.clone(), self.inner.replay_bytes).await?;
+            // The ring the previous daemon held died with it, so an adopted
+            // session would otherwise replay nothing until tmux's own screen
+            // capture — which is one screenful, not scrollback (#171).
+            let bridge = PtyBridge::start_restoring(
+                fifo_path.clone(),
+                self.inner.replay_bytes,
+                self.restored_scrollback(id),
+            )
+            .await?;
             // Re-pipe so output starts flowing again. `Tmux::pipe_pane` runs
             // without tmux's `-o` flag, so any stale pipe the previous daemon
             // left running to a now-unread FIFO is replaced (NAP-228).
@@ -894,6 +1052,8 @@ impl SessionManager {
                     naming: false,
                     focused_since: None,
                     awaiting_suppressed: false,
+                    resuming: false,
+                    flushed_bytes: 0,
                 },
             );
             let _ = self
@@ -912,7 +1072,9 @@ impl SessionManager {
     ///
     /// A missing stored value (first run) or an unreadable current one
     /// (non-Linux, restricted `/proc`, unwired in this build) both read as
-    /// "no reboot", which is today's behaviour: fall back to probing tmux.
+    /// "no reboot". Either way the tmux probe runs (#174): the answer only
+    /// tells recovery whether a tmux server that did not answer is coming
+    /// back, never which sessions are alive.
     async fn check_boot_id(&self) -> bool {
         let Some(cfg) = &self.inner.boot_id else {
             return false;
@@ -959,7 +1121,8 @@ impl SessionManager {
     /// `rebooted` (#164) only changes what gets logged: a detected reboot
     /// names the count even when it is zero, so the operator sees that
     /// recovery told a reboot from a restart instead of a log that stays
-    /// silent either way.
+    /// silent either way. It never decides membership — everything here got
+    /// here by tmux not holding it (#174).
     async fn register_dormant(&self, rebooted: bool) {
         let stored = match self.inner.sidecar.load_all() {
             Ok(stored) => stored,
@@ -970,6 +1133,7 @@ impl SessionManager {
         };
         let mut registered = Vec::new();
         let mut pruned = 0usize;
+        let mut discarded = 0usize;
         {
             let mut sessions = self.inner.sessions.lock().await;
             for mut session in stored {
@@ -981,9 +1145,21 @@ impl SessionManager {
                     if let Err(e) = self.inner.sidecar.delete(session.id) {
                         tracing::warn!(error = ?e, id = %session.id, "recover: failed to prune sidecar with missing cwd");
                     }
+                    self.prune_scrollback(session.id);
                     pruned += 1;
                     continue;
                 }
+                if self.inner.resume_policy == ResumeOnBoot::Never {
+                    // #167: this host wants a reboot to be a clean slate, so
+                    // the sidecar goes the way #163's does — nothing is
+                    // registered and the next recovery finds nothing.
+                    if let Err(e) = self.inner.sidecar.delete(session.id) {
+                        tracing::warn!(error = ?e, id = %session.id, "recover: failed to prune sidecar under resume_on_boot=never");
+                    }
+                    self.prune_scrollback(session.id);
+                    discarded += 1;
+                    continue;
+                }
                 let id = session.id;
                 session.status = SessionStatus::Dormant;
                 session.foreground = None;
@@ -1006,6 +1182,8 @@ impl SessionManager {
                         naming: false,
                         focused_since: None,
                         awaiting_suppressed: false,
+                        resuming: false,
+                        flushed_bytes: 0,
                     },
                 );
                 registered.push(session);
@@ -1025,12 +1203,45 @@ impl SessionManager {
         if pruned > 0 {
             tracing::info!(pruned, "pruned sidecars whose cwd no longer exists");
         }
+        let resumable: Vec<Uuid> = registered.iter().map(|s| s.id).collect();
         for session in registered {
             let _ = self
                 .inner
                 .events
                 .send(CatalogEvent::SessionAdded { session });
         }
+
+        // #167: say which policy ran and how many sessions it acted on, so a
+        // host that resumes agents unattended says so in its own log. The
+        // match is exhaustive on purpose: a fourth policy must decide what it
+        // does here rather than inherit whichever arm happened to be last.
+        match self.inner.resume_policy {
+            ResumeOnBoot::Dormant => tracing::info!(
+                policy = ResumeOnBoot::Dormant.as_str(),
+                sessions = resumable.len(),
+                "resume_on_boot: sessions left dormant for the operator"
+            ),
+            ResumeOnBoot::Never => tracing::info!(
+                policy = ResumeOnBoot::Never.as_str(),
+                sessions = discarded,
+                "resume_on_boot: pruned the sidecars of sessions with no live pane"
+            ),
+            ResumeOnBoot::Auto => {
+                tracing::info!(
+                    policy = ResumeOnBoot::Auto.as_str(),
+                    sessions = resumable.len(),
+                    "resume_on_boot: resuming sessions with no live pane"
+                );
+                let presets = self.inner.resume_presets.clone();
+                for id in resumable {
+                    // The same call the client's POST /api/sessions/:id/resume
+                    // makes — one resume path, not two.
+                    if let Err(e) = self.resume(id, &presets).await {
+                        tracing::warn!(error = %e, id = %id, "resume_on_boot=auto: resume failed; session stays dormant");
+                    }
+                }
+            }
+        }
     }
 
     /// Forward keystrokes to a session's pane. A dead or dormant session is
@@ -1193,10 +1404,12 @@ impl SessionManager {
     /// and stays silent on the event stream at steady state.
     ///
     /// `agent_session_id` (#161) is the agent's own conversation id, parsed
-    /// by the caller from the hook payload's `session_id` field. It is
-    /// written once — the first report that carries one wins — and the
-    /// sidecar is only rewritten on that transition, never on every report,
-    /// since an agent's session id never changes mid-session.
+    /// by the caller from the hook payload's `session_id` field. The
+    /// sidecar is rewritten only when the reported id differs from the
+    /// stored one — a repeated report of the same id is pure churn from the
+    /// hook's `PreToolUse` cadence and writes nothing, but a resume (#176)
+    /// issues a new conversation id and must replace the stale one, or
+    /// every later resume reattaches the pre-resume conversation.
     pub async fn report_activity(
         &self,
         id: Uuid,
@@ -1218,16 +1431,16 @@ impl SessionManager {
                 state,
                 self.inner.report_ttl,
             ));
-            let newly_identified =
-                rec.session.agent_session_id.is_none() && agent_session_id.is_some();
-            if newly_identified {
+            let id_changed =
+                agent_session_id.is_some() && rec.session.agent_session_id != agent_session_id;
+            if id_changed {
                 rec.session.agent_session_id = agent_session_id;
                 self.inner
                     .sidecar
                     .write(&rec.session)
                     .map_err(SessionError::Sidecar)?;
             }
-            (changed || newly_identified).then(|| rec.session.clone())
+            (changed || id_changed).then(|| rec.session.clone())
         };
         if let Some(session) = updated {
             let _ = self
@@ -1610,6 +1823,36 @@ async fn run_blocked_work_pass(inner: &Arc<Inner>) {
     }
 }
 
+/// Mirror every live bridge's scrollback to disk. See
+/// [`SessionManager::flush_scrollback`].
+async fn run_scrollback_pass(inner: &Arc<Inner>) {
+    let Some(store) = inner.scrollback.clone() else {
+        return;
+    };
+    // Snapshot under the lock, write with it released: the writes are
+    // filesystem work and every attach needs this map.
+    let pending: Vec<(Uuid, Vec<u8>)> = {
+        let mut sessions = inner.sessions.lock().await;
+        sessions
+            .iter_mut()
+            .filter_map(|(id, rec)| {
+                let bridge = rec.bridge.as_ref()?;
+                let seen = bridge.bytes_seen();
+                if seen == rec.flushed_bytes {
+                    return None;
+                }
+                rec.flushed_bytes = seen;
+                Some((*id, bridge.scrollback()))
+            })
+            .collect()
+    };
+    for (id, bytes) in pending {
+        if let Err(e) = store.write(id, &bytes) {
+            tracing::warn!(error = ?e, %id, "scrollback: flush failed");
+        }
+    }
+}
+
 /// Mark a session whose tmux session has disappeared as [`SessionStatus::Dead`]
 /// and settle its worktree per the natural-exit rule (#82).
 ///
@@ -1636,6 +1879,15 @@ async fn mark_session_dead(inner: &Arc<Inner>, id: Uuid) {
                 // attach.
                 if let Some(bridge) = record.bridge.as_mut() {
                     bridge.seal();
+                    // The clean-exit half of #171: the last thing the session
+                    // printed is what a later resume should open on, and it
+                    // will produce nothing more for the flush pass to notice.
+                    if let Some(store) = &inner.scrollback {
+                        if let Err(e) = store.write(id, &bridge.scrollback()) {
+                            tracing::warn!(error = ?e, %id, "scrollback: final flush failed");
+                        }
+                        record.flushed_bytes = bridge.bytes_seen();
+                    }
                 }
                 let session = record.session.clone();
                 let worktree = record
@@ -1757,6 +2009,14 @@ async fn cleanup_record(
     if let Err(e) = inner.sidecar.delete(id) {
         tracing::warn!(error = ?e, %id, "cleanup: failed to delete sidecar");
     }
+    // #171: the session is gone, so its scrollback goes with it — nothing
+    // will ever read the file again, and leaving it would keep a killed
+    // session's output on disk indefinitely.
+    if let Some(store) = &inner.scrollback {
+        if let Err(e) = store.delete(id) {
+            tracing::warn!(error = ?e, %id, "cleanup: failed to delete scrollback");
+        }
+    }
     if let (Some(base), Some(path)) = (&record.worktree_base_repo, &record.worktree_path) {
         if keep_worktree {
             tracing::info!(

crates/maestro-daemon/tests/activity_report.rs

@@ -187,10 +187,30 @@ async fn a_reported_session_id_is_recorded_on_the_session() {
     assert_eq!(s.agent_session_id.as_deref(), Some("conv-abc123"));
 }
 
-/// #161: a session id does not change mid-session, so once set it is never
-/// overwritten by a later report — even one that names a different id.
+/// #161: repeated reports of the same id are pure churn from the hook's
+/// `PreToolUse` cadence — the sidecar must not be rewritten for them.
 #[tokio::test]
-async fn a_reported_session_id_is_written_once_and_never_replaced() {
+async fn a_repeated_report_of_the_same_session_id_writes_nothing() {
+    let fx = fixture(Duration::from_secs(300)).await;
+    let id = spawn_sleeping(&fx.manager).await;
+
+    fx.manager
+        .report_activity(id, ActivityState::Active, None, Some("conv-first".into()))
+        .await
+        .unwrap();
+    fx.manager
+        .report_activity(id, ActivityState::Active, None, Some("conv-first".into()))
+        .await
+        .unwrap();
+
+    let s = observed(&fx.manager, id).await;
+    assert_eq!(s.agent_session_id.as_deref(), Some("conv-first"));
+}
+
+/// #176: a resume issues a new conversation id. The stored id must move to
+/// it, or every later resume reattaches the stale pre-resume conversation.
+#[tokio::test]
+async fn a_reported_session_id_is_replaced_when_it_differs() {
     let fx = fixture(Duration::from_secs(300)).await;
     let id = spawn_sleeping(&fx.manager).await;
 
@@ -204,7 +224,14 @@ async fn a_reported_session_id_is_written_once_and_never_replaced() {
         .unwrap();
 
     let s = observed(&fx.manager, id).await;
-    assert_eq!(s.agent_session_id.as_deref(), Some("conv-first"));
+    assert_eq!(s.agent_session_id.as_deref(), Some("conv-second"));
+
+    let stored = fx
+        .sidecar
+        .read(id)
+        .expect("sidecar read")
+        .expect("sidecar present");
+    assert_eq!(stored.agent_session_id.as_deref(), Some("conv-second"));
 }
 
 #[tokio::test]

crates/maestro-daemon/tests/common/mod.rs

@@ -151,27 +151,40 @@ pub struct DaemonFixture {
 
 impl DaemonFixture {
     pub fn new() -> Self {
-        Self::build(None)
+        Self::build(None, maestro_daemon::pty::DEFAULT_REPLAY_BYTES)
+    }
+
+    /// Variant whose manager runs a custom PTY replay budget (NAP-229), which
+    /// also caps the scrollback it persists (#171).
+    pub fn new_with_replay(replay_bytes: usize) -> Self {
+        Self::build(None, replay_bytes)
     }
 
     /// Variant whose manager has a durable spawn history (#24) attached, with
     /// `retention_days` recency cap, logging to `<state_dir>/spawn-history.jsonl`.
     pub fn new_with_history(retention_days: u32) -> Self {
-        Self::build(Some(retention_days))
+        Self::build(
+            Some(retention_days),
+            maestro_daemon::pty::DEFAULT_REPLAY_BYTES,
+        )
     }
 
-    fn build(history_days: Option<u32>) -> Self {
+    fn build(history_days: Option<u32>, replay_bytes: usize) -> Self {
         let tmux = TmuxFixture::new();
         let state_dir = TempDir::new().unwrap();
         let sidecar = SidecarStore::new(state_dir.path().join("sessions"));
         // Default placement: worktrees land under their base repo (#56).
         let worktree_mgr = WorktreeManager::new(None);
-        let mut manager = SessionManager::new(
+        let mut manager = SessionManager::new_with_replay(
             Arc::new(tmux.tmux.clone()),
             sidecar,
             worktree_mgr,
             state_dir.path().join("fifos"),
+            replay_bytes,
         );
+        // #171: production wires this, so the fixtures do too — every test
+        // then exercises the same flush/restore/delete paths the daemon runs.
+        manager = manager.with_scrollback(state_dir.path().join("scrollback"));
         if let Some(days) = history_days {
             manager =
                 manager.with_spawn_history(SpawnHistory::new(Self::history_path(&state_dir), days));
@@ -198,4 +211,8 @@ impl DaemonFixture {
     pub fn fifo_dir(&self) -> PathBuf {
         self.state_dir.path().join("fifos")
     }
+
+    pub fn scrollback_dir(&self) -> PathBuf {
+        self.state_dir.path().join("scrollback")
+    }
 }

crates/maestro-daemon/tests/resume.rs

@@ -357,3 +357,237 @@ async fn post_resume_on_an_unknown_session_is_404() {
         .unwrap();
     assert_eq!(resp.status(), reqwest::StatusCode::NOT_FOUND);
 }
+
+// ---------------------------------------------------------------------------
+// Races (#175)
+// ---------------------------------------------------------------------------
+
+/// Real tmux, with the *first* `new_session_piped` held inside the call until
+/// the test releases it. Later calls pass straight through, so a second resume
+/// racing the first never blocks on a gate the first one owns.
+///
+/// This is what makes the two races below deterministic rather than timing
+/// dependent: the resume under test is parked in exactly the window the bug
+/// lives in, and the racing operation runs to completion before it moves.
+#[derive(Debug)]
+struct GatedSpawn {
+    inner: maestro_daemon::tmux::Tmux,
+    entered: tokio::sync::mpsc::UnboundedSender<()>,
+    release: std::sync::Mutex<std::sync::mpsc::Receiver<()>>,
+    gated: std::sync::atomic::AtomicBool,
+}
+
+/// The test's end of a [`GatedSpawn`].
+struct Gate {
+    entered: tokio::sync::mpsc::UnboundedReceiver<()>,
+    release: std::sync::mpsc::SyncSender<()>,
+}
+
+impl Gate {
+    /// Wait until the gated spawn is parked mid-resume.
+    async fn wait_entered(&mut self) {
+        tokio::time::timeout(Duration::from_secs(10), self.entered.recv())
+            .await
+            .expect("resume reached the tmux spawn")
+            .expect("gate alive");
+    }
+
+    fn release(&self) {
+        self.release.send(()).expect("gate receiver alive");
+    }
+}
+
+fn gated_panes(inner: maestro_daemon::tmux::Tmux) -> (Arc<GatedSpawn>, Gate) {
+    let (entered_tx, entered_rx) = tokio::sync::mpsc::unbounded_channel();
+    let (release_tx, release_rx) = std::sync::mpsc::sync_channel(1);
+    (
+        Arc::new(GatedSpawn {
+            inner,
+            entered: entered_tx,
+            release: std::sync::Mutex::new(release_rx),
+            gated: std::sync::atomic::AtomicBool::new(true),
+        }),
+        Gate {
+            entered: entered_rx,
+            release: release_tx,
+        },
+    )
+}
+
+impl maestro_daemon::panes::Panes for GatedSpawn {
+    fn new_session_piped(
+        &self,
+        id: Uuid,
+        cwd: &std::path::Path,
+        command: &str,
+        session_env: &std::collections::BTreeMap<String, String>,
+        sink_command: &str,
+    ) -> Result<(), maestro_daemon::tmux::TmuxError> {
+        if self.gated.swap(false, std::sync::atomic::Ordering::SeqCst) {
+            let _ = self.entered.send(());
+            let _ = self.release.lock().expect("gate mutex").recv();
+        }
+        self.inner
+            .new_session_piped(id, cwd, command, session_env, sink_command)
+    }
+
+    fn kill_session(&self, id: Uuid) -> Result<(), maestro_daemon::tmux::TmuxError> {
+        self.inner.kill_session(id)
+    }
+
+    fn list_sessions(&self) -> Result<Vec<Uuid>, maestro_daemon::tmux::TmuxError> {
+        self.inner.list_sessions()
+    }
+
+    fn pane_info(
+        &self,
+        id: Uuid,
+    ) -> Result<Option<maestro_daemon::tmux::PaneInfo>, maestro_daemon::tmux::TmuxError> {
+        self.inner.pane_info(id)
+    }
+
+    fn send_keys(&self, id: Uuid, bytes: &[u8]) -> Result<(), maestro_daemon::tmux::TmuxError> {
+        self.inner.send_keys(id, bytes)
+    }
+
+    fn pipe_pane(&self, id: Uuid, sink: &str) -> Result<(), maestro_daemon::tmux::TmuxError> {
+        self.inner.pipe_pane(id, sink)
+    }
+
+    fn resize_window(
+        &self,
+        id: Uuid,
+        cols: u16,
+        rows: u16,
+    ) -> Result<(), maestro_daemon::tmux::TmuxError> {
+        self.inner.resize_window(id, cols, rows)
+    }
+
+    fn capture_pane(&self, id: Uuid) -> Result<Option<Vec<u8>>, maestro_daemon::tmux::TmuxError> {
+        self.inner.capture_pane(id)
+    }
+}
+
+/// A dormant session whose manager's tmux spawn can be parked by the test.
+async fn dormant_gated(fx: &DaemonFixture, cwd: &TempDir) -> (SessionManager, Session, Gate) {
+    let session = fx
+        .manager
+        .create(CreateSessionRequest {
+            cwd: cwd.path().to_path_buf(),
+            command: "sleep 30".into(),
+            env: HashMap::new(),
+            metadata: HashMap::new(),
+            worktree: None,
+        })
+        .await
+        .expect("create");
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+
+    let (panes, gate) = gated_panes(fx.tmux.tmux.clone());
+    let manager = SessionManager::new(
+        panes,
+        maestro_daemon::sidecar::SidecarStore::new(fx.sidecar_dir()),
+        maestro_daemon::worktree::WorktreeManager::new(None),
+        fx.fifo_dir(),
+    );
+    manager.recover().await.expect("recover");
+    assert_eq!(
+        manager.status(session.id).await,
+        Some(SessionStatus::Dormant)
+    );
+    (manager, session, gate)
+}
+
+/// A kill that lands while a resume is inside the tmux spawn must not leave
+/// the sidecar behind. The kill deletes it; the resume writes it back a moment
+/// later and then discovers its own record is gone. The sidecar it resurrected
+/// would otherwise make the next recovery register a ghost dormant session for
+/// a session that was killed.
+#[tokio::test]
+async fn a_kill_during_resume_leaves_no_sidecar_behind() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let (manager, session, mut gate) = dormant_gated(&fx, &cwd).await;
+    let id = session.id;
+
+    let resuming = tokio::spawn({
+        let manager = manager.clone();
+        async move { manager.resume(id, &[]).await }
+    });
+    gate.wait_entered().await;
+
+    manager.kill(id, false).await.expect("kill mid-resume");
+    gate.release();
+
+    let err = resuming.await.unwrap().expect_err("record is gone");
+    assert!(matches!(err, SessionError::NotFound(got) if got == id));
+
+    assert!(
+        maestro_daemon::sidecar::SidecarStore::new(fx.sidecar_dir())
+            .read(id)
+            .expect("read sidecar")
+            .is_none(),
+        "the killed session's sidecar must not be resurrected by the losing resume"
+    );
+
+    // And so the next recovery registers no ghost.
+    let after = revive(&fx);
+    after.recover().await.expect("recover");
+    assert!(
+        after.snapshot().await.is_empty(),
+        "a killed session must not come back dormant"
+    );
+}
+
+/// Two resumes of the same session must not both proceed. The second is
+/// refused on the state — `already_running` — not on its tmux spawn colliding
+/// with the first one's.
+#[tokio::test]
+async fn a_second_resume_is_refused_while_the_first_is_in_flight() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let (manager, session, mut gate) = dormant_gated(&fx, &cwd).await;
+    let id = session.id;
+
+    let resuming = tokio::spawn({
+        let manager = manager.clone();
+        async move { manager.resume(id, &[]).await }
+    });
+    gate.wait_entered().await;
+
+    let err = manager
+        .resume(id, &[])
+        .await
+        .expect_err("a resume is already in flight");
+    assert!(matches!(err, SessionError::AlreadyRunning(got) if got == id));
+    assert_eq!(err.to_error_info().code, "already_running");
+
+    gate.release();
+    let first = resuming.await.unwrap().expect("the first resume wins");
+    assert_eq!(first.status, SessionStatus::Running);
+    assert_eq!(
+        fx.tmux.tmux.list_sessions().expect("list"),
+        vec![id],
+        "exactly one pane for the session"
+    );
+}
+
+/// A refused resume must not leave the session marked in-flight: once the
+/// reason for the refusal is gone, resuming works.
+#[tokio::test]
+async fn a_refused_resume_can_be_retried() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let (manager, session) = dormant_session(&fx, &cwd, "sleep 30", "retried").await;
+
+    std::fs::remove_dir_all(cwd.path()).unwrap();
+    let err = manager
+        .resume(session.id, &[])
+        .await
+        .expect_err("cwd is gone");
+    assert_eq!(err.to_error_info().code, "cwd_missing");
+
+    std::fs::create_dir_all(cwd.path()).unwrap();
+    let resumed = manager.resume(session.id, &[]).await.expect("resume");
+    assert_eq!(resumed.status, SessionStatus::Running);
+}

crates/maestro-daemon/tests/resume_policy.rs

@@ -0,0 +1,161 @@
+//! `[daemon].resume_on_boot` (#167): what recovery does with sessions whose
+//! tmux pane is gone but whose sidecar survived, against real tmux.
+//!
+//! The default is deliberately the timid one. Resuming agents unattended
+//! seconds after a reboot is a foot-gun, so `auto` is opted into per host and
+//! `never` throws the sidecars away instead.
+
+mod common;
+
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+
+use common::{wait_for_async, DaemonFixture};
+use maestro_config::ResumeOnBoot;
+use maestro_daemon::SessionManager;
+use maestro_protocol::{CreateSessionRequest, Session, SessionStatus, SpawnPreset};
+use tempfile::TempDir;
+
+/// A fresh manager over the same tmux server, sidecars and fifos — what the
+/// daemon builds after a restart — carrying `policy`.
+fn revive(fx: &DaemonFixture, policy: ResumeOnBoot, presets: Vec<SpawnPreset>) -> SessionManager {
+    SessionManager::new(
+        Arc::new(fx.tmux.tmux.clone()),
+        maestro_daemon::sidecar::SidecarStore::new(fx.sidecar_dir()),
+        maestro_daemon::worktree::WorktreeManager::new(None),
+        fx.fifo_dir(),
+    )
+    .with_resume_policy(policy, presets)
+}
+
+/// A preset whose `resume_command` touches a marker file, so a test can see
+/// that the resumed session ran the resume command rather than the original.
+fn marker_preset(command: &str, marker_dir: &TempDir) -> Vec<SpawnPreset> {
+    vec![SpawnPreset {
+        id: "agent".into(),
+        label: "Agent".into(),
+        command: command.into(),
+        resume_command: Some(format!(
+            "sh -c 'touch {}/resumed; sleep 30'",
+            marker_dir.path().display()
+        )),
+    }]
+}
+
+async fn create(fx: &DaemonFixture, cwd: &TempDir, command: &str) -> Session {
+    fx.manager
+        .create(CreateSessionRequest {
+            cwd: cwd.path().to_path_buf(),
+            command: command.into(),
+            env: HashMap::new(),
+            metadata: HashMap::new(),
+            worktree: None,
+        })
+        .await
+        .expect("create")
+}
+
+/// The default: a reboot leaves the session listed but not running, waiting
+/// for the operator.
+#[tokio::test]
+async fn dormant_is_the_default_and_leaves_the_session_waiting() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let session = create(&fx, &cwd, "sleep 30").await;
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+
+    let revived = revive(&fx, ResumeOnBoot::Dormant, Vec::new());
+    revived.recover().await.expect("recover");
+
+    assert_eq!(
+        revived.status(session.id).await,
+        Some(SessionStatus::Dormant)
+    );
+    assert!(fx
+        .sidecar_dir()
+        .join(format!("{}.json", session.id))
+        .exists());
+    assert!(
+        fx.tmux.tmux.list_sessions().expect("list").is_empty(),
+        "nothing was respawned"
+    );
+}
+
+/// `auto` brings each one back itself, down the same path a client's resume
+/// takes: same uuid, same sidecar, the preset's `resume_command`.
+#[tokio::test]
+async fn auto_resumes_every_dormant_session_during_recovery() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let markers = TempDir::new().unwrap();
+    let session = create(&fx, &cwd, "sleep 30").await;
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+
+    let revived = revive(&fx, ResumeOnBoot::Auto, marker_preset("sleep 30", &markers));
+    revived.recover().await.expect("recover");
+
+    assert_eq!(
+        revived.status(session.id).await,
+        Some(SessionStatus::Running),
+        "auto resumed it"
+    );
+    assert_eq!(
+        fx.tmux.tmux.list_sessions().expect("list"),
+        vec![session.id],
+        "back under the same uuid"
+    );
+    let stored = maestro_daemon::sidecar::SidecarStore::new(fx.sidecar_dir())
+        .read(session.id)
+        .expect("read sidecar")
+        .expect("same sidecar");
+    assert_eq!(stored.created_at, session.created_at);
+    assert_eq!(stored.command, "sleep 30", "the original command is kept");
+
+    // It really ran the preset's resume command, not the spawn command.
+    let marker = markers.path().join("resumed");
+    wait_for_async(Duration::from_secs(5), || {
+        let marker = marker.clone();
+        async move { marker.exists() }
+    })
+    .await;
+    assert!(marker.exists(), "resume_command ran");
+}
+
+/// `never` makes a reboot a clean slate: the sidecars go, and the session is
+/// not listed at all.
+#[tokio::test]
+async fn never_prunes_the_sidecars_instead() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let session = create(&fx, &cwd, "sleep 30").await;
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+
+    let revived = revive(&fx, ResumeOnBoot::Never, Vec::new());
+    revived.recover().await.expect("recover");
+
+    assert!(!revived.has(session.id).await, "not listed");
+    assert!(
+        !fx.sidecar_dir()
+            .join(format!("{}.json", session.id))
+            .exists(),
+        "sidecar pruned"
+    );
+    assert!(fx.tmux.tmux.list_sessions().expect("list").is_empty());
+}
+
+/// The policy only decides what happens to sessions that lost their pane. A
+/// session tmux still holds is adopted running by the first pass under every
+/// policy — including `never`, which must not delete a live session's sidecar.
+#[tokio::test]
+async fn the_policy_leaves_sessions_tmux_still_holds_alone() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let live = create(&fx, &cwd, "sleep 30").await;
+
+    let revived = revive(&fx, ResumeOnBoot::Never, Vec::new());
+    assert_eq!(revived.recover().await.expect("recover"), 1);
+
+    assert_eq!(revived.status(live.id).await, Some(SessionStatus::Running));
+    assert!(fx.sidecar_dir().join(format!("{}.json", live.id)).exists());
+}

crates/maestro-daemon/tests/scrollback.rs

@@ -0,0 +1,219 @@
+//! Persisted PTY scrollback across a reboot (#171), against real tmux.
+//!
+//! Everything else in the reboot milestone restores what a session *is*; this
+//! restores what it *looked like*. The cases that matter are the ones a clean
+//! shutdown never gets to run: the host went down, and the only scrollback
+//! that exists is whatever the last periodic flush wrote.
+
+mod common;
+
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::time::Duration;
+
+use common::{wait_for_async, DaemonFixture};
+use maestro_daemon::scrollback::ScrollbackStore;
+use maestro_daemon::SessionManager;
+use maestro_protocol::{CreateSessionRequest, Session, SessionStatus};
+use tempfile::TempDir;
+use uuid::Uuid;
+
+/// A fresh manager over the same tmux server, sidecars, fifos and scrollback —
+/// what the daemon builds after a restart.
+fn revive(fx: &DaemonFixture) -> SessionManager {
+    SessionManager::new(
+        Arc::new(fx.tmux.tmux.clone()),
+        maestro_daemon::sidecar::SidecarStore::new(fx.sidecar_dir()),
+        maestro_daemon::worktree::WorktreeManager::new(None),
+        fx.fifo_dir(),
+    )
+    .with_scrollback(fx.scrollback_dir())
+}
+
+async fn spawn(fx: &DaemonFixture, cwd: &TempDir, command: &str) -> Session {
+    fx.manager
+        .create(CreateSessionRequest {
+            cwd: cwd.path().to_path_buf(),
+            command: command.into(),
+            env: HashMap::new(),
+            metadata: HashMap::new(),
+            worktree: None,
+        })
+        .await
+        .expect("create")
+}
+
+/// Wait until the session has printed `needle`, flushing as we go — the same
+/// thing the liveness watcher's flush pass does on its timer.
+async fn wait_for_flushed(
+    manager: &SessionManager,
+    store: &ScrollbackStore,
+    id: Uuid,
+    needle: &str,
+) {
+    wait_for_async(Duration::from_secs(10), || async {
+        manager.flush_scrollback().await;
+        String::from_utf8_lossy(&store.read(id, 64 * 1024)).contains(needle)
+    })
+    .await;
+}
+
+fn replay_text(replay: &[u8]) -> String {
+    String::from_utf8_lossy(replay).to_string()
+}
+
+/// The keystone: the host reboots with no exit path run at all, and the
+/// resumed session opens showing what it printed before.
+#[tokio::test]
+async fn a_resumed_session_replays_the_scrollback_from_before_the_reboot() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let store = ScrollbackStore::new(fx.scrollback_dir());
+
+    let session = spawn(&fx, &cwd, "sh -c 'echo before-the-reboot; sleep 30'").await;
+    wait_for_flushed(&fx.manager, &store, session.id, "before-the-reboot").await;
+
+    // The reboot: tmux dies where it stands. No kill, no seal, no shutdown
+    // hook — only whatever the periodic flush had already written.
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+    let manager = revive(&fx);
+    manager.recover().await.expect("recover");
+    assert_eq!(
+        manager.status(session.id).await,
+        Some(SessionStatus::Dormant)
+    );
+
+    manager.resume(session.id, &[]).await.expect("resume");
+    let (replay, _rx) = manager
+        .subscribe_pty_with_replay(session.id)
+        .await
+        .expect("attachable again");
+    assert!(
+        replay_text(&replay).contains("before-the-reboot"),
+        "resumed pane should open on the old scrollback, got {:?}",
+        replay_text(&replay)
+    );
+}
+
+/// A daemon restart leaves tmux alone, so the session is adopted rather than
+/// resumed — and it too lost its in-memory ring with the old process.
+#[tokio::test]
+async fn an_adopted_session_replays_the_scrollback_of_the_previous_daemon() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let store = ScrollbackStore::new(fx.scrollback_dir());
+
+    let session = spawn(&fx, &cwd, "sh -c 'echo printed-earlier; sleep 30'").await;
+    wait_for_flushed(&fx.manager, &store, session.id, "printed-earlier").await;
+
+    // Daemon restart: the manager goes, tmux stays.
+    let manager = revive(&fx);
+    manager.recover().await.expect("recover");
+    assert_eq!(
+        manager.status(session.id).await,
+        Some(SessionStatus::Running)
+    );
+
+    let (replay, _rx) = manager
+        .subscribe_pty_with_replay(session.id)
+        .await
+        .expect("attachable");
+    assert!(
+        replay_text(&replay).contains("printed-earlier"),
+        "adopted pane should carry the old scrollback ahead of tmux's screen, got {:?}",
+        replay_text(&replay)
+    );
+}
+
+/// Killing a session takes its scrollback with it: nothing will ever read the
+/// file again, and the operator's output should not linger on disk.
+#[tokio::test]
+async fn killing_a_session_deletes_its_scrollback() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let store = ScrollbackStore::new(fx.scrollback_dir());
+
+    let session = spawn(&fx, &cwd, "sh -c 'echo doomed; sleep 30'").await;
+    wait_for_flushed(&fx.manager, &store, session.id, "doomed").await;
+    assert!(store.path_for(session.id).exists());
+
+    fx.manager.kill(session.id, false).await.expect("kill");
+    assert!(
+        !store.path_for(session.id).exists(),
+        "scrollback should not outlive the session"
+    );
+}
+
+/// A session that exits on its own does get an exit path, and takes it: the
+/// last thing it printed is flushed when the bridge is sealed, with no wait
+/// for the next tick.
+#[tokio::test]
+async fn a_session_that_exits_flushes_its_last_output() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let store = ScrollbackStore::new(fx.scrollback_dir());
+
+    let session = spawn(&fx, &cwd, "sh -c 'echo last-words'").await;
+    fx.manager
+        .start_liveness_watcher(Duration::from_millis(100));
+    wait_for_async(Duration::from_secs(10), || async {
+        fx.manager.status(session.id).await == Some(SessionStatus::Dead)
+    })
+    .await;
+
+    assert!(
+        String::from_utf8_lossy(&store.read(session.id, 64 * 1024)).contains("last-words"),
+        "a sealed bridge should have flushed what it printed"
+    );
+}
+
+/// Nothing here may fail a resume. A scrollback file that is gone, or one full
+/// of bytes no terminal would emit, degrades to a pane with no history — not
+/// to an error.
+#[tokio::test]
+async fn a_missing_or_corrupt_scrollback_still_resumes() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let store = ScrollbackStore::new(fx.scrollback_dir());
+
+    let gone = spawn(&fx, &cwd, "sleep 30").await;
+    let garbled = spawn(&fx, &cwd, "sleep 30").await;
+    fx.manager.flush_scrollback().await;
+
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+    std::fs::create_dir_all(store.dir()).expect("scrollback dir");
+    std::fs::remove_file(store.path_for(gone.id)).ok();
+    std::fs::write(store.path_for(garbled.id), [0xff, 0xfe, 0x00, 0x80]).expect("write garbage");
+
+    let manager = revive(&fx);
+    manager.recover().await.expect("recover");
+    for id in [gone.id, garbled.id] {
+        manager.resume(id, &[]).await.expect("resume");
+        assert_eq!(manager.status(id).await, Some(SessionStatus::Running));
+        assert!(manager.subscribe_pty_with_replay(id).await.is_some());
+    }
+}
+
+/// The disk cost is the replay budget, not a second one of its own.
+#[tokio::test]
+async fn the_flushed_file_stays_inside_the_replay_budget() {
+    let fx = DaemonFixture::new_with_replay(256);
+    let cwd = TempDir::new().unwrap();
+    let store = ScrollbackStore::new(fx.scrollback_dir());
+
+    let session = spawn(
+        &fx,
+        &cwd,
+        "sh -c 'for i in $(seq 1 400); do echo line-$i; done; sleep 30'",
+    )
+    .await;
+    wait_for_flushed(&fx.manager, &store, session.id, "line-400").await;
+
+    let size = std::fs::metadata(store.path_for(session.id))
+        .expect("stat scrollback")
+        .len();
+    assert!(
+        size <= 256,
+        "scrollback file was {size} bytes, budget is 256"
+    );
+}

crates/maestro-daemon/tests/session_manager.rs

@@ -326,12 +326,14 @@ fn revive_with_boot_id(
     )
 }
 
-/// #164: a changed boot id means every session from the previous boot is
-/// already dead, so recovery must not trust tmux to say so — it skips the
-/// probe outright. Proven by leaving the tmux session alive: if the probe
-/// ran, it would come back `Running`, not `Dormant`.
+/// #174: a boot id can change without the tmux server dying — a restored
+/// state directory, a VM or container snapshot restore, a truncated
+/// `boot_id` file. `tmux list-sessions` stays the single source of truth for
+/// liveness, so a session tmux still holds is adopted `Running` no matter
+/// what the boot id says. Registering it dormant instead stranded it: resume
+/// would collide with the tmux name still in use and fail forever.
 #[tokio::test]
-async fn recover_skips_tmux_probe_when_boot_id_changed() {
+async fn recover_adopts_live_session_when_boot_id_changed() {
     let fx = DaemonFixture::new();
     let cwd = TempDir::new().unwrap();
     let s = fx
@@ -349,15 +351,15 @@ async fn recover_skips_tmux_probe_when_boot_id_changed() {
     let boot_id_path = fx.state_dir.path().join("boot_id");
     std::fs::write(&boot_id_path, "previous-boot").unwrap();
 
-    // tmux session is still alive — a probe would find it if one ran.
+    // tmux session is still alive: the probe runs and finds it.
     let revived = revive_with_boot_id(&fx, boot_id_path.clone(), Some("current-boot"));
     let adopted = revived.recover().await.expect("recover");
-    assert_eq!(adopted, 0, "reboot detected: nothing adopted live");
+    assert_eq!(adopted, 1, "tmux still holds the session; adopt it");
 
     let snap = revived.snapshot().await;
     assert_eq!(snap.len(), 1);
     assert_eq!(snap[0].id, s.id);
-    assert_eq!(snap[0].status, SessionStatus::Dormant);
+    assert_eq!(snap[0].status, SessionStatus::Running);
 
     // The current boot id is now the stored one, for the next comparison.
     assert_eq!(
@@ -366,6 +368,38 @@ async fn recover_skips_tmux_probe_when_boot_id_changed() {
     );
 }
 
+/// #164, still true after #174: a real reboot takes the tmux server with it,
+/// so the probe comes back empty and every sidecar is registered dormant.
+#[tokio::test]
+async fn recover_registers_dormant_when_boot_id_changed_and_tmux_gone() {
+    let fx = DaemonFixture::new();
+    let cwd = TempDir::new().unwrap();
+    let s = fx
+        .manager
+        .create(CreateSessionRequest {
+            cwd: cwd.path().to_path_buf(),
+            command: "sleep 30".into(),
+            env: HashMap::new(),
+            metadata: HashMap::new(),
+            worktree: None,
+        })
+        .await
+        .expect("create");
+
+    let boot_id_path = fx.state_dir.path().join("boot_id");
+    std::fs::write(&boot_id_path, "previous-boot").unwrap();
+    fx.tmux.tmux.kill_server().expect("kill tmux server");
+
+    let revived = revive_with_boot_id(&fx, boot_id_path, Some("current-boot"));
+    let adopted = revived.recover().await.expect("recover");
+    assert_eq!(adopted, 0, "no tmux server survived the reboot");
+
+    let snap = revived.snapshot().await;
+    assert_eq!(snap.len(), 1);
+    assert_eq!(snap[0].id, s.id);
+    assert_eq!(snap[0].status, SessionStatus::Dormant);
+}
+
 /// #164: an unchanged boot id is a plain restart — recovery probes tmux as
 /// before and adopts the still-live session `Running`.
 #[tokio::test]

docs/INSTALL.md

@@ -243,6 +243,26 @@ Pick `/tmp` if the daemon must outlive your login session and you can't run
 tmux sockets with it). The env var `MAESTRO_DAEMON_TMUX_TMPDIR` takes
 precedence over the file.
 
+## What happens to sessions after a reboot
+
+A host reboot kills the tmux server, so every session the daemon knew is dead
+while its record on disk survives. What recovery does with those is
+`[daemon].resume_on_boot`:
+
+```toml
+[daemon]
+resume_on_boot = "dormant"   # "dormant" (default) | "auto" | "never"
+```
+
+- `dormant` — the sessions are listed but not running, and you resume the ones
+  you want from the web UI or the TUI.
+- `auto` — the daemon resumes them all itself during startup. Only pick this
+  on a host where restarting every agent unattended is safe: they come back on
+  the flags they were spawned with, with nobody watching.
+- `never` — the records are deleted and a reboot is a clean slate.
+
+See `docs/RESTART.md` for the full contract.
+
 ## What the systemd unit looks like
 
 The wizard writes a *user* unit (`systemctl --user`), since `maestro daemon`

docs/RESTART.md

@@ -75,7 +75,13 @@ A daemon restart leaves tmux alone, so recovery re-adopts every
 beyond the reconnect above.
 
 A **host reboot** is different: it kills the tmux server, so there is nothing
-to adopt while every session's sidecar is still on disk. Recovery makes a
+to adopt while every session's sidecar is still on disk. Recovery always asks
+tmux — `tmux list-sessions` is the single source of truth for liveness, and the
+kernel's `boot_id` only tells a reboot from a restart in the log (#164, #174).
+A boot id changes without the tmux server dying whenever the state directory is
+restored, a VM or container snapshot is rolled back, or the `boot_id` file is
+truncated; a session tmux still holds is adopted `Running` in every one of
+those cases. Recovery makes a
 second pass over those sidecars and registers each as a `Dormant` session
 (#162) — id, name, cwd, command, worktree and `created_at` intact. A dormant
 session has no pane and no PTY bridge: it cannot be attached to, typed into or
@@ -92,10 +98,74 @@ conversation id is respawned on the command it was created with. `Session::comma
 always keeps that original command, so a second resume resolves the same preset.
 Resume is refused with `409` for a session that is already running
 (`already_running`) and for one whose `cwd` has disappeared (`cwd_missing`) —
-neither is something a respawn could fix. The other thing to do with a dormant
+neither is something a respawn could fix. A session with a resume already in
+flight is refused as `already_running` too (#175): the respawn happens with
+the session map unlocked, so the record is marked in-flight before it starts
+and the second request is answered on that state rather than on its tmux spawn
+colliding with the first one's. A resume that loses a race with a kill takes
+its own sidecar back off disk, so the killed session does not return as a
+dormant ghost on the next recovery. The other thing to do with a dormant
 session is dismiss it, which removes the sidecar so the next recovery does not
 bring it back.
 
+## What the pane looks like when it comes back
+
+The PTY replay ring lives in memory, so a resumed session would otherwise open
+on a blank pane: the conversation is back, the terminal history the operator
+was reading is not. Each bridge's ring is mirrored to
+`<state_dir>/scrollback/<id>.bin` (#171) and handed to the new bridge on
+resume, and on the recovery pass that re-adopts a session tmux still holds —
+that one lost its ring with the previous daemon too. The restored bytes go out
+ahead of the live ones, and are dropped once the ring holds a full budget of
+new output of its own, which is the point where everything restored is older
+than the budget covers.
+
+The file is capped at the same `[daemon] pty_replay_bytes` budget the ring
+uses — one budget, not two — and it is deleted when the session is killed, and
+when recovery prunes its sidecar. Nothing here can fail a resume: a missing,
+unreadable or oversized file degrades to an empty pane, which is exactly what
+the operator had before it existed.
+
+It is flushed every 10 seconds by the liveness watcher, and once more when a
+session's bridge is sealed on exit. The timer is the part that matters: a host
+reboot runs no exit path at all, so scrollback written only on a clean shutdown
+would be missing in the one case this exists for. Ten seconds bounds the loss
+to the last few lines — scrollback is visual continuity, not state — and costs
+at most one file write per *active* session per tick, since a session whose
+bridge has read nothing since the last flush is skipped.
+
+## What recovery does with them: `resume_on_boot`
+
+Whether a recovered-but-dead session is left dormant at all is the operator's
+call, per host, via `[daemon].resume_on_boot` (#167):
+
+| value       | what recovery does                                                    |
+| ----------- | --------------------------------------------------------------------- |
+| `"dormant"` | **Default.** Register each one dormant and wait for the operator.      |
+| `"auto"`    | Resume each one during recovery, itself.                               |
+| `"never"`   | Prune the sidecars instead — a reboot is a clean slate.                |
+
+`dormant` is the default because resuming agents unattended is a foot-gun:
+they come back on whatever flags they were spawned with, seconds after a
+reboot, with nobody watching. `auto` exists for hosts where that is known to
+be safe, and it is opted into rather than inherited.
+
+`auto` calls the same `resume` the client's `POST /api/sessions/:id/resume`
+does — same uuid, same sidecar, the preset's `resume_command` — so there is
+one resume path and not two. A session that fails to resume is logged and
+stays dormant; the rest still come back.
+
+`never` prunes the same way #163 does for a missing `cwd`: the sidecar is
+deleted and nothing is registered, so the session is not listed and the next
+recovery does not find it. Whatever worktree it owned is left on disk.
+
+The policy only decides the fate of sessions with no live pane. A session
+tmux still holds is adopted `Running` by the first pass under every policy,
+including `never`.
+
+Recovery logs the policy that ran and the number of sessions it acted on, so
+a host that resumes agents unattended says so in its own journal.
+
 Attaching to a session that has no bridge answers `409 not_attachable`, not
 `404 unknown_session`: the session exists, and a client that cannot tell "gone"
 from "not running" cannot offer the resume.