hive-agent: self-announce boot instead of host-side kick_agent

docs/turn-loop/claude-invocation.md

@@ -222,11 +222,19 @@ to ack), plus an inline `({unread} more pending — drain via …)` hint
 when `unread > 0`. Claude drives any further `recv`/`send` itself via
 the embedded MCP server.
 
-Whenever hive-c0re starts / restarts / rebuilds a container, it
+`hive-agent` self-announces once on every boot:
-also drops a `system` message into the agent's inbox via
+`boot_announce::announce` pushes a one-line "you were just
-`Coordinator::kick_agent` — a one-line "you were just (re)started,
+(re)started, check /state/ for your notes, your session is intact"
-check /state/ for your notes, your session is intact." The
+todo the moment the in-agent todo store opens, before the serve loop
-next turn picks it up like any other inbox message.
+starts. This replaced `hive-c0re`'s old `Coordinator::kick_agent`
+calls at the tail of the rebuild and start DAG nodes — those fired a
+broker message per DAG node, so one rebuild (always a real
+`StopForUpdate` → `Swap` → `RebuildBookkeeping` → `Reconcile` →
+`Start` cycle) queued two separate restart messages for a single
+actual boot. The harness itself is the one side that knows for
+certain whether a boot happened, so it dedups for free: exactly one
+todo per real process start, surfaced the normal todo-wake way
+rather than as an inbox message.
 
 ## On-boot files
 

hive-agent/src/boot_announce.rs

@@ -0,0 +1,47 @@
+//! One-shot startup self-announce: pushes a todo on every harness boot so
+//! the agent gets a "you were just (re)started" notice without the host
+//! having to send one.
+//!
+//! Replaces `hive-c0re`'s old `Coordinator::kick_agent` calls at the tail
+//! of the rebuild and start DAG nodes, which an operator reported as
+//! duplicate restart messages: those fired one broker message per DAG
+//! node, so a single rebuild — always a real
+//! `StopForUpdate` → `Swap` → `RebuildBookkeeping` → `Reconcile` → `Start`
+//! cycle — queued *two* separate restart messages for one actual boot, and
+//! an agent paused for days could rack up dozens more from repeated
+//! host-side sweeps. The harness itself has perfect knowledge of "did a
+//! boot actually happen": it only runs this once, at the top of
+//! `serve_main`. Self-announcing here dedups for free — one todo per real
+//! process start, no host-side bookkeeping needed.
+//!
+//! No first-spawn-vs-restart distinction: every boot gets the same notice
+//! (the old first-spawn message didn't reliably fire anyway, so unifying
+//! the two loses nothing).
+//!
+//! Keyless on purpose: each boot is a genuinely new occurrence, not a
+//! reconciled condition like `disk_watch`'s, so there's nothing to dedup
+//! *against* within one store — `Todos::upsert`'s keyless path always
+//! inserts a fresh row and always reports `changed`, which is exactly
+//! "wake me, this happened" with no extra bookkeeping.
+
+use tokio::sync::Notify;
+
+use crate::todos::Todos;
+
+/// Producing subsystem marker for the boot-announce todo row.
+const SUBSYSTEM: &str = "harness";
+
+/// Push the "you were just (re)started" todo and wake the turn loop.
+/// Best-effort: a store error is logged, never fatal to boot.
+pub fn announce(label: &str, todos: &Todos, wake: &Notify) {
+    let summary = format!(
+        "You were just (re)started. If you were mid-task, check \
+         `/agents/{label}/state/` for your notes and pick up where you left \
+         off — claude's `--continue` session is intact, so prior context is \
+         still in your window."
+    );
+    match todos.upsert(SUBSYSTEM, None, &summary, None, false) {
+        Ok(_) => wake.notify_one(),
+        Err(e) => tracing::warn!(error = ?e, "boot_announce: todo upsert failed"),
+    }
+}

hive-agent/src/main.rs

@@ -9,6 +9,7 @@
 //! Single bin crate: the module tree below (formerly this crate's `lib.rs`,
 //! before lib + bin were collapsed into one) plus the serve loop.
 
+mod boot_announce;
 mod claude_md_watch;
 mod claude_usage_watch;
 mod db_migrate;
@@ -613,6 +614,12 @@ async fn serve_main<S: Surface>(socket: &Path, poll_ms: u64) -> Result<()> {
     };
     tokio::spawn(reminder_timer::run(reminder_store.clone(), local_msg_tx));
     let (todo_wake, todos_store) = spawn_todo_socket(reminder_store.clone(), &bus);
+    // Self-announce this boot as a todo instead of relying on hive-c0re's
+    // old `kick_agent` broker messages — see `boot_announce`'s module doc
+    // for why this is strictly better than a host-side kick.
+    if let Some(store) = &todos_store {
+        boot_announce::announce(&label, store, &todo_wake);
+    }
     if matches!(initial, LoginState::NeedsLogin) {
         login::wait_for_login(
             &claude_dir,

hive-c0re/src/job_queue/exec.rs

@@ -513,10 +513,16 @@ async fn run_rebuild_bookkeeping(coord: &Arc<Coordinator>, name: &str) -> Result
     // access all recover without a hive-c0re restart. (No matrix step: the
     // swarm mints an agent's matrix account, not this hive.)
     crate::forge::sync_agent(name, crate::forge::core_token().as_deref()).await;
-    // Wake the agent on its next turn so claude sees a "you were rebuilt"
+    // No `kick_agent` here any more — reporting from `hive-agent`'s own
-    // hint; rescan so dashboards drop the "needs update" chip; lock bump →
+    // boot instead avoids the duplicate-restart-message problem this used
+    // to cause: the container is still stopped at this point (`swap_update` requires it stopped and
+    // never restarts it — the DAG's tail `Reconcile`/`Start` does), so a
+    // host-side "you were rebuilt" broker message would just queue
+    // alongside `run_start`'s own kick for the same boot. `hive-agent` now
+    // self-announces once on its own startup instead — one message per
+    // actual boot, not one per DAG node that happens to run during it.
+    // Rescan so dashboards drop the "needs update" chip; lock bump →
     // meta-inputs re-render.
-    coord.kick_agent(name, "container rebuilt");
     coord.rescan_containers_and_emit().await;
     crate::dashboard::emit_meta_inputs_snapshot(coord);
     Ok(())
@@ -674,7 +680,9 @@ async fn run_start(coord: &Arc<Coordinator>, name: &str) -> Result<()> {
     // (no background poll) — c0re owns the listener lifecycle, so
     // register here rather than waiting for a sweep.
     coord.register_agent(name)?;
-    coord.kick_agent(name, "container started");
+    // No `kick_agent` here any more — `hive-agent` pushes its own "you
+    // were just (re)started" todo on boot now, so this node doesn't need
+    // to tell it. See the matching comment in `run_rebuild_bookkeeping`.
     coord.rescan_containers_and_emit().await;
     Ok(())
 }