The agent's status is the CLI's word, not whether a process runs

Desktop said "running" as soon as `lembas service run` was started, while the CLI was still
dialling, waiting for a terminal that holds the link, or refused by the server -- the reason went
only to agent.log. Now the supervisor reads what the CLI itself writes: `service.json` in its
state directory (connecting, linked, waiting + why, stopped + why) for the pid it started, the
line it printed before that, and its last line when it ends. A service the app did not start is
described the same way.

A run the CLI ends with 0 because trying again will not help (a refused token, no login) is no
longer restarted every second: it waits five minutes, or until linking again starts it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src-tauri/src/agent.rs

@@ -4,6 +4,12 @@
 //! the web UI run here, in this computer's projects. It is the CLI's own service loop, unchanged;
 //! the app only starts it, restarts it when it fails, and stops it (SIGTERM first, so a turn in
 //! progress ends cleanly) when the agent is switched off or the app quits.
+//!
+//! What it shows is the CLI's own word: a process that runs is not an agent that is linked. While
+//! it runs, the state is what the service writes to its `service.json` (connecting, linked,
+//! waiting and why, stopped and why); before that, the line it printed; when it ends, its last
+//! line. A service the CLI says cannot go on (a refused token, no login) is not restarted every
+//! second: it waits until linking again starts it.
 
 use std::fs::{File, OpenOptions};
 use std::path::PathBuf;
@@ -20,6 +26,8 @@ use crate::cli;
 const RESTART_FOR_UPDATE: i32 = 75;
 const BACKOFF: Duration = Duration::from_secs(10);
 const STOP_GRACE: Duration = Duration::from_secs(15);
+/// How long to wait before trying again after the CLI said trying again will not help.
+const REFUSED_WAIT: Duration = Duration::from_secs(300);
 
 #[derive(Debug, Clone, PartialEq, Serialize)]
 #[serde(tag = "state", content = "detail", rename_all = "snake_case")]
@@ -28,25 +36,83 @@ pub enum State {
     Off,
     /// This computer is not signed in to the server yet.
     NotLinked,
+    /// The process is started; the CLI has not said anything of its link yet.
     Starting,
+    /// The CLI is dialling the server.
+    Connecting,
+    /// The CLI says it is linked: agent chats can use this computer.
     Running,
-    /// Another `lembas service` on this computer holds the link (its pid).
+    /// Not linked right now, and trying again: the CLI's own reason (a terminal holds the link,
-    Elsewhere(u32),
+    /// the server cannot be reached, the link dropped).
-    /// It ended with an error; trying again shortly.
+    Waiting(String),
+    /// The CLI will not try again by itself: the server refused this computer's token, or the
+    /// login is missing. Linking again is the way out.
+    Refused(String),
+    /// Another `lembas service` on this computer holds the link.
+    Elsewhere(Elsewhere),
+    /// It ended with an error; trying again shortly. The CLI's last words, where it said any.
     Failed(String),
 }
 
+/// A `lembas service` this app did not start, and what it says of its link.
+#[derive(Debug, Clone, PartialEq, Serialize)]
+pub struct Elsewhere {
+    pub pid: u32,
+    /// It says it is linked.
+    pub linked: bool,
+    /// Why not, when it is not (its own words; empty while it has said nothing).
+    pub why: String,
+}
+
 impl State {
     pub fn label(&self) -> String {
         match self {
             State::Off => "Agent: off".into(),
             State::NotLinked => "Agent: this computer is not linked".into(),
             State::Starting => "Agent: starting…".into(),
+            State::Connecting => "Agent: connecting…".into(),
             State::Running => "Agent: running".into(),
-            State::Elsewhere(pid) => format!("Agent: running as a service (pid {pid})"),
+            State::Waiting(why) => format!("Agent: waiting — {why}"),
+            State::Refused(why) => format!("Agent: refused — {why}"),
+            State::Elsewhere(e) if e.linked => {
+                format!("Agent: running as a service (pid {})", e.pid)
+            }
+            State::Elsewhere(e) => format!("Agent: a service (pid {}) is not linked", e.pid),
             State::Failed(why) => format!("Agent: stopped — {why}"),
         }
     }
+
+    /// Whether this is a state the CLI's own status file decides, while the process runs.
+    fn from_cli(&self) -> bool {
+        matches!(
+            self,
+            State::Starting | State::Connecting | State::Running | State::Waiting(_)
+        )
+    }
+}
+
+/// What the CLI's status says, as a state of the app's. `before`: the state shown now, so that a
+/// new dial after a failed one keeps the reason it is waiting (the CLI writes `connecting` at
+/// every try, `waiting` between them) instead of flickering between the two.
+pub fn from_link(link: &cli::Link, before: &State) -> State {
+    let why = link.detail.clone().unwrap_or_default();
+    match link.state.as_str() {
+        "linked" => State::Running,
+        "connecting" => match before {
+            State::Waiting(w) if link.attempts > 1 => State::Waiting(w.clone()),
+            _ => State::Connecting,
+        },
+        "waiting" => State::Waiting(why),
+        "stopped" => State::Refused(why),
+        _ => State::Starting,
+    }
+}
+
+/// What a running process has said before its status file says anything: the one line the
+/// CLI's service writes while a terminal on this computer holds the link for `/remote`.
+fn from_output(line: &str) -> Option<State> {
+    line.contains("holds the link")
+        .then(|| State::Waiting(line.trim().to_string()))
 }
 
 type Notify = Arc<dyn Fn(&State) + Send + Sync>;
@@ -124,15 +190,33 @@ impl Agent {
                 continue;
             }
             if let Some(pid) = cli::running_service().filter(|p| Some(*p) != self.child_pid()) {
-                if !self.set(generation, State::Elsewhere(pid)) {
+                // Not ours, but this computer's agent all the same: say what it says.
+                let link = cli::link_of(pid);
+                let linked = link.as_ref().is_some_and(|l| l.state == "linked");
+                let why = link
+                    .map(|l| match l.detail {
+                        Some(d) if !d.is_empty() => d,
+                        _ => l.state,
+                    })
+                    .unwrap_or_default();
+                if !self.set(
+                    generation,
+                    State::Elsewhere(Elsewhere {
+                        pid,
+                        linked,
+                        why: if linked { String::new() } else { why },
+                    }),
+                ) {
                     return;
                 }
-                self.pause(generation, BACKOFF);
+                self.pause(generation, Duration::from_secs(3));
                 continue;
             }
             if !self.set(generation, State::Starting) {
                 return;
             }
+            // Where this run's output starts in the log: its last line is its reason.
+            let from = std::fs::metadata(&self.log).map(|m| m.len()).unwrap_or(0);
             let child = match self.spawn() {
                 Ok(c) => c,
                 Err(e) => {
@@ -146,6 +230,7 @@ impl Agent {
                     continue;
                 }
             };
+            let pid = child.id();
             {
                 let mut g = self.inner.lock().unwrap();
                 if g.generation != generation {
@@ -155,23 +240,39 @@ impl Agent {
                 }
                 g.child = Some(child);
             }
-            self.set(generation, State::Running);
+            let code = self.wait_child(generation, pid, from);
-            let code = self.wait_child(generation);
             let mut g = self.inner.lock().unwrap();
             if g.generation != generation {
                 return;
             }
             g.child = None;
             drop(g);
+            let said = last_line(&self.log, from);
             match code {
                 Some(RESTART_FOR_UPDATE) => continue,
                 Some(0) => {
-                    // It ended by itself without an error: nothing to do but start it again.
+                    // It ended by itself without an error: the CLI's word for "starting again
-                    self.pause(generation, Duration::from_secs(1));
+                    // will not help" -- the server refused the token, or there is no login.
+                    // Waiting a long while (linking again starts it at once).
+                    let why = cli::link_of(pid)
+                        .filter(|l| l.state == "stopped")
+                        .and_then(|l| l.detail)
+                        .or_else(|| said.map(|s| strip_prefix(&s)));
+                    match why {
+                        Some(why) => {
+                            if !self.set(generation, State::Refused(why)) {
+                                return;
+                            }
+                            self.pause(generation, REFUSED_WAIT);
+                        }
+                        None => self.pause(generation, Duration::from_secs(1)),
+                    }
                 }
                 other => {
-                    let why = other.map_or("ended by a signal".to_string(), |c| {
+                    let why = said.map(|s| strip_prefix(&s)).unwrap_or_else(|| {
+                        other.map_or("ended by a signal".to_string(), |c| {
                             format!("exit status {c}")
+                        })
                     });
                     if !self.set(generation, State::Failed(why)) {
                         return;
@@ -217,7 +318,11 @@ impl Agent {
     }
 
     /// Poll the child until it exits (its exit code), or until this supervisor is replaced.
-    fn wait_child(&self, generation: u64) -> Option<i32> {
+    /// Meanwhile the state shown is what the CLI says of its link: its status file once it has
+    /// written one, before that the line it printed (a terminal holding the link).
+    fn wait_child(&self, generation: u64, pid: u32, from: u64) -> Option<i32> {
+        // Every other poll (once a second) the status is read again.
+        let mut read = true;
         loop {
             {
                 let mut g = self.inner.lock().unwrap();
@@ -230,6 +335,21 @@ impl Agent {
                     Some(Err(_)) | None => return None,
                 }
             }
+            if read {
+                let now = self.state();
+                if now.from_cli() {
+                    let next = match cli::link_of(pid) {
+                        Some(link) => from_link(&link, &now),
+                        None => last_line(&self.log, from)
+                            .and_then(|l| from_output(&l))
+                            .unwrap_or(State::Starting),
+                    };
+                    if !self.set(generation, next) {
+                        return None;
+                    }
+                }
+            }
+            read = !read;
             thread::sleep(Duration::from_millis(500));
         }
     }
@@ -283,3 +403,99 @@ fn terminate(mut child: Child) {
     let _ = child.kill();
     let _ = child.wait();
 }
+
+/// The last line the log holds after `from` (this run's output), if it wrote any. Only the end of
+/// the file is read.
+fn last_line(log: &std::path::Path, from: u64) -> Option<String> {
+    use std::io::{Read, Seek, SeekFrom};
+    let mut f = File::open(log).ok()?;
+    let len = f.metadata().ok()?.len();
+    let start = from.max(len.saturating_sub(4096));
+    if start >= len {
+        return None;
+    }
+    f.seek(SeekFrom::Start(start)).ok()?;
+    let mut buf = Vec::new();
+    f.take(len - start).read_to_end(&mut buf).ok()?;
+    String::from_utf8_lossy(&buf)
+        .lines()
+        .rev()
+        .map(str::trim)
+        .find(|l| !l.is_empty())
+        .map(String::from)
+}
+
+/// `lembas service: not logged in` → `not logged in`: the CLI's error line without its name.
+fn strip_prefix(line: &str) -> String {
+    line.strip_prefix("lembas service: ")
+        .unwrap_or(line)
+        .to_string()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn link(state: &str, detail: Option<&str>, attempts: u32) -> cli::Link {
+        cli::Link {
+            state: state.into(),
+            detail: detail.map(String::from),
+            pid: 1,
+            attempts,
+        }
+    }
+
+    #[test]
+    fn the_state_shown_is_the_clis_word_for_its_link() {
+        let s = State::Starting;
+        assert_eq!(from_link(&link("linked", None, 1), &s), State::Running);
+        assert_eq!(
+            from_link(&link("connecting", None, 1), &s),
+            State::Connecting
+        );
+        let waiting = from_link(
+            &link("waiting", Some("could not reach the instance"), 1),
+            &s,
+        );
+        assert_eq!(
+            waiting,
+            State::Waiting("could not reach the instance".into())
+        );
+        // The next dial keeps the reason it is waiting, rather than flicker.
+        assert_eq!(from_link(&link("connecting", None, 2), &waiting), waiting);
+        assert_eq!(
+            from_link(&link("stopped", Some("refused"), 2), &waiting),
+            State::Refused("refused".into())
+        );
+        // Not running is never shown as running.
+        for state in ["connecting", "waiting", "stopped", "?"] {
+            assert_ne!(from_link(&link(state, None, 1), &s), State::Running);
+        }
+    }
+
+    #[test]
+    fn a_terminal_holding_the_link_is_said() {
+        assert_eq!(
+            from_output("a terminal here holds the link (/remote); waiting for it to close"),
+            Some(State::Waiting(
+                "a terminal here holds the link (/remote); waiting for it to close".into()
+            ))
+        );
+        assert_eq!(from_output("linked"), None);
+    }
+
+    #[test]
+    fn the_last_line_is_this_runs_own() {
+        let p = std::env::temp_dir().join(format!("llembas-agent-log-{}", std::process::id()));
+        std::fs::write(&p, "old run\nlinked\n").unwrap();
+        let from = std::fs::metadata(&p).unwrap().len();
+        assert_eq!(last_line(&p, from), None);
+        let mut f = OpenOptions::new().append(true).open(&p).unwrap();
+        std::io::Write::write_all(&mut f, b"connecting\nlembas service: not logged in\n\n")
+            .unwrap();
+        let said = last_line(&p, from).unwrap();
+        assert_eq!(said, "lembas service: not logged in");
+        assert_eq!(strip_prefix(&said), "not logged in");
+        std::fs::remove_file(&p).unwrap();
+    }
+}

src-tauri/src/cli.rs

@@ -8,6 +8,8 @@
 use std::path::{Path, PathBuf};
 use std::sync::OnceLock;
 
+use serde::Deserialize;
+
 static BUNDLED: OnceLock<PathBuf> = OnceLock::new();
 
 /// Where the bundle put the CLI: the app's resource directory (`/usr/lib/llembas-desktop/` from
@@ -108,6 +110,32 @@ fn running_in(state: &Path) -> Option<u32> {
     (pid != 0 && alive(pid)).then_some(pid)
 }
 
+/// What a running `lembas service run` says of its link to the server, as it writes it to
+/// `service.json` in its state directory at every change: `connecting`, `linked`, `waiting`
+/// (not linked now, trying again; `detail` says why) or `stopped` (it will not try again;
+/// `detail` says why -- the server refused the token, or speaks another protocol). `pid` is the
+/// process that wrote it: a file left by an earlier run is not this one's word.
+#[derive(Clone, Debug, PartialEq, Deserialize)]
+pub struct Link {
+    pub state: String,
+    #[serde(default)]
+    pub detail: Option<String>,
+    #[serde(default)]
+    pub pid: u32,
+    #[serde(default)]
+    pub attempts: u32,
+}
+
+/// The link status the CLI's service wrote last, if it was `pid` that wrote it.
+pub fn link_of(pid: u32) -> Option<Link> {
+    link_in(&state_dir()).filter(|l| l.pid == pid)
+}
+
+fn link_in(state: &Path) -> Option<Link> {
+    let text = std::fs::read_to_string(state.join("service.json")).ok()?;
+    serde_json::from_str(&text).ok()
+}
+
 /// This computer's host name: what the CLI sent the server as `device_host` at login
 /// (`os.hostname()`, the kernel's node name), so the server can tell which device is this one.
 #[cfg(unix)]
@@ -166,6 +194,33 @@ mod tests {
         fs::remove_dir_all(&d).unwrap();
     }
 
+    #[test]
+    fn the_link_status_is_read_as_the_service_writes_it() {
+        let d = scratch("link");
+        assert_eq!(link_in(&d), None);
+        // As `lembas service run` writes it (src/service.ts): LinkStatus and its pid.
+        fs::write(
+            d.join("service.json"),
+            r#"{"state":"waiting","instance":"https://ai.example","since":"2026-01-01T00:00:00.000Z","detail":"could not reach the instance; trying again in 4s","attempts":3,"pid":4242}"#,
+        )
+        .unwrap();
+        let l = link_in(&d).unwrap();
+        assert_eq!(l.state, "waiting");
+        assert_eq!(l.pid, 4242);
+        assert_eq!(l.attempts, 3);
+        assert!(l.detail.unwrap().starts_with("could not reach"));
+        // `linked` carries no detail.
+        fs::write(
+            d.join("service.json"),
+            r#"{"state":"linked","instance":"x","since":"y","attempts":1,"pid":7}"#,
+        )
+        .unwrap();
+        assert_eq!(link_in(&d).unwrap().detail, None);
+        fs::write(d.join("service.json"), "not json").unwrap();
+        assert_eq!(link_in(&d), None);
+        fs::remove_dir_all(&d).unwrap();
+    }
+
     #[cfg(unix)]
     #[test]
     fn a_lock_counts_only_while_its_process_lives() {

ui/js/settings.js

@@ -9,8 +9,10 @@ const AGENT = {
   off: () => [t("The agent is off."), ""],
   not_linked: () => [t("This computer is not linked to the server yet."), ""],
   starting: () => [t("The agent is starting…"), ""],
+  connecting: () => [t("The agent is connecting to the server…"), ""],
   running: () => [t("The agent is running: agent chats can use this computer."), "is-ok"],
-  elsewhere: (pid) => [t("A LLeMbas CLI service already runs here (pid %(pid)s); it is this computer's agent.", { pid }), "is-ok"],
+  waiting: (why) => [t("The agent is not connected: %(why)s. It keeps trying.", { why }), "is-error"],
+  refused: (why) => [t("The agent will not connect: %(why)s. Link this computer again.", { why }), "is-error"],
   failed: (why) => [t("The agent stopped: %(why)s. Trying again shortly.", { why }), "is-error"],
 }
 
@@ -18,10 +20,18 @@ let folderError = ""
 let linking = null // { user_code, uri } while the computer is being linked
 let linkError = ""
 
+/** The agent's state in a sentence, and its class: the CLI's word for its link. */
 export function agentWords(s) {
+  const a = s.agent
   const noFolders = s.remote && !(s.remote.enabled && s.remote.roots.length)
-  if (s.agent.state === "running" && noFolders) return [t("The agent is connected, but may open no folder yet: add one below."), "is-error"]
+  if (a.state === "elsewhere") {
-  return (AGENT[s.agent.state] ?? AGENT.off)(s.agent.detail)
+    const pid = a.detail?.pid
+    if (!a.detail?.linked) return [t("A LLeMbas CLI service already runs here (pid %(pid)s), but is not connected: %(why)s", { pid, why: a.detail?.why || "…" }), "is-error"]
+    if (noFolders) return [t("The agent is connected, but may open no folder yet: add one below."), "is-error"]
+    return [t("A LLeMbas CLI service already runs here (pid %(pid)s); it is this computer's agent.", { pid }), "is-ok"]
+  }
+  if (a.state === "running" && noFolders) return [t("The agent is connected, but may open no folder yet: add one below."), "is-error"]
+  return (AGENT[a.state] ?? AGENT.off)(a.detail || "…")
 }
 
 function section(id, title, ...body) {

ui/js/titlebar.js

@@ -9,12 +9,35 @@ const WORDS = {
   off: () => t("agent off"),
   not_linked: () => t("not linked"),
   starting: () => t("agent starting…"),
+  connecting: () => t("agent connecting…"),
   running: () => t("agent running"),
+  waiting: () => t("agent waiting"),
+  refused: () => t("agent refused"),
   elsewhere: () => t("agent running (service)"),
+  elsewhere_waiting: () => t("agent waiting"),
   failed: () => t("agent stopped"),
   no_folders: () => t("agent: no folders"),
 }
 
+/**
+ * The pill's state: what the CLI says of its link, not whether a process runs. Linked but allowed
+ * into no folder is said too -- a green light would say it works when it cannot.
+ */
+export function pillState(s) {
+  const a = s.agent
+  const noFolders = s.remote && !(s.remote.enabled && s.remote.roots.length)
+  const linked = a.state === "running" || (a.state === "elsewhere" && a.detail?.linked)
+  if (linked && noFolders) return "no_folders"
+  if (a.state === "elsewhere" && !a.detail?.linked) return "elsewhere_waiting"
+  return a.state
+}
+
+/** Why the agent is not working, in the CLI's words, when it said. */
+export function agentWhy(a) {
+  if (a.state === "elsewhere") return a.detail?.why || ""
+  return typeof a.detail === "string" ? a.detail : ""
+}
+
 export function renderBar(s) {
   let host = ""
   try {
@@ -22,12 +45,11 @@ export function renderBar(s) {
   } catch {}
   $("host").textContent = host
   $("host").title = s.server || ""
-  // Connected but allowed into no folder: a green light would say it works when it cannot.
+  const state = pillState(s)
-  const noFolders = s.remote && !(s.remote.enabled && s.remote.roots.length)
+  const reason = agentWhy(s.agent)
-  const state = s.agent.state === "running" && noFolders ? "no_folders" : s.agent.state
   $("agent").dataset.state = state
   $("agent-text").textContent = WORDS[state] ? WORDS[state]() : state
-  $("agent").title = state === "failed" ? t("Agent stopped: %(why)s", { why: s.agent.detail }) : t("Agent")
+  $("agent").title = reason ? t("Agent: %(why)s", { why: reason }) : t("Agent")
   $("update").hidden = !s.update
   $("update").textContent = s.update ? t("update %(version)s", { version: s.update }) : t("update")
   $("update").title = s.update ? t("LLeMbas %(version)s is available (this is %(current)s)", { version: s.update, current: s.version }) : ""