harness todos: ack reconciled todos instead of deleting them

hive-agent/src/disk_watch.rs

@@ -15,7 +15,10 @@
 //! loose-ends list instead of waking the agent every tick. Only crossing
 //! into a new bucket (or a change in which directories are big enough to
 //! list) speaks up again. Dropping back under the threshold clears the
-//! todo.
+//! todo. Acknowledging the todo (`cancel_loose_end`) does not defeat this —
+//! `Todos::mark_done` acks the row rather than deleting it, so the next
+//! probe still has the prior summary to compare against instead of seeing
+//! an empty table and re-announcing an unchanged condition as new.
 //!
 //! Scoped by design: the `statvfs` reads the *whole filesystem*, which on
 //! a shared host volume (several agents' state dirs on the same

hive-agent/src/todos.rs

@@ -15,9 +15,18 @@
 //! same item is idempotent, and a producer can list / clear / rebuild only
 //! its own set (e.g. matrix wipes + recreates its todos on daemon restart).
 //!
-//! Removal has two paths (mara's call): the producing subsystem `clear`s a
-//! todo it has resolved (keyed by subsystem + key), or the agent itself
-//! `mark_done`s one by id. Both delete the row.
+//! Removal/dismissal has two different paths. The **producing subsystem**
+//! `clear`s a todo it has itself resolved (keyed by subsystem + key) — a
+//! one-shot event has nothing left to represent once resolved, so this
+//! hard-deletes the row. The **agent** `mark_done`s one by id (via
+//! `cancel_loose_end`) — that only means "stop showing it to me", so it sets
+//! `acked` instead of deleting: a *reconciled* producer like `disk_watch`
+//! (re-derives its summary on a timer, not a one-shot event) needs the row
+//! to survive ack so [`Todos::upsert`] can still tell "unchanged" from
+//! "genuinely different". Acked rows stay out of [`Todos::list`] but keep
+//! anchoring that comparison; a materially different summary un-acks +
+//! wakes again. Acked rows are reaped after a retention window
+//! (`vacuum.rs`) so the table doesn't grow unbounded.
 
 use std::path::Path;
 use std::sync::Mutex;
@@ -26,6 +35,10 @@ use anyhow::{Context, Result};
 use hive_sh4re::wire_time::now_unix;
 use rusqlite::{Connection, params};
 
+/// SQL bootstrap. `CREATE TABLE IF NOT EXISTS` so first-boot agents and
+/// existing ones converge on the same base shape; additive columns land via
+/// [`MIGRATIONS`] as try-and-ignore `ALTER TABLE`s so existing dbs catch up
+/// without a hard schema-version bump.
 const SCHEMA: &str = r"
 CREATE TABLE IF NOT EXISTS todos (
     id            INTEGER PRIMARY KEY AUTOINCREMENT,
@@ -42,6 +55,17 @@ CREATE UNIQUE INDEX IF NOT EXISTS idx_todos_dedup
     ON todos (subsystem, subsystem_key);
 ";
 
+/// Additive column migrations — see `turn_stats.rs`'s identical pattern.
+/// Each runs unconditionally and ignores `duplicate column name` errors, so
+/// re-running on an already-migrated db is a silent no-op.
+const MIGRATIONS: &[&str] = &[
+    "ALTER TABLE todos ADD COLUMN acked INTEGER NOT NULL DEFAULT 0",
+    // No default: only meaningful once `acked = 1`; NULL on every
+    // pre-migration / never-acked row, which `reap_acked`'s `acked_at < ?`
+    // comparison naturally excludes (NULL compares false, not true).
+    "ALTER TABLE todos ADD COLUMN acked_at INTEGER",
+];
+
 /// One dynamic, subsystem-pushed todo. Timestamps are unix seconds; the
 /// consumer derives `age_seconds` from `updated_at`.
 #[derive(Debug, Clone)]
@@ -75,6 +99,18 @@ impl Todos {
         let conn =
             Connection::open(path).with_context(|| format!("open todos db {}", path.display()))?;
         conn.execute_batch(SCHEMA).context("apply todos schema")?;
+        for stmt in MIGRATIONS {
+            // Ignore "duplicate column name" — the migration already ran.
+            // Any other error is logged but doesn't fail open(): the base
+            // schema still works and we'd rather keep the harness alive
+            // than crash on an upgrade hiccup.
+            if let Err(e) = conn.execute(stmt, []) {
+                let msg = e.to_string();
+                if !msg.contains("duplicate column name") {
+                    tracing::warn!(error = %msg, stmt, "todos migration failed");
+                }
+            }
+        }
         Ok(Self {
             conn: Mutex::new(conn),
         })
@@ -84,10 +120,14 @@ impl Todos {
     /// when `key` is `Some` and already present. A `None` key never
     /// conflicts, so it always inserts a fresh row.
     ///
-    /// Returns `(id, changed)` where `changed` is `true` when the row is
-    /// new OR its `summary`/`source` actually differed — the caller uses
-    /// this to decide whether to signal the turn loop (re-pushing an
-    /// identical keyed todo is a no-op and must not re-wake).
+    /// Returns `(id, changed)` where `changed` is `true` when the row is new
+    /// OR its `summary`/`source` actually differed — the caller uses this to
+    /// decide whether to signal the turn loop (re-pushing an identical keyed
+    /// todo is a no-op and must not re-wake). This holds **regardless of
+    /// whether the existing row is acked**: an identical re-push of an
+    /// acked row stays quiet (and stays acked/hidden), while a genuinely
+    /// different summary un-acks it and reports `changed = true` — the
+    /// agent's earlier dismissal doesn't suppress a real re-deterioration.
     ///
     /// # Errors
     ///
@@ -121,8 +161,12 @@ impl Todos {
             if unchanged {
                 return Ok((id, false));
             }
+            // A materially different summary clears any prior ack — this is
+            // new information, not a re-announcement of what was dismissed.
             conn.execute(
-                "UPDATE todos SET summary = ?1, source = ?2, updated_at = ?3 WHERE id = ?4",
+                "UPDATE todos SET summary = ?1, source = ?2, updated_at = ?3, \
+                        acked = 0, acked_at = NULL \
+                 WHERE id = ?4",
                 params![summary, source, now, id],
             )?;
             return Ok((id, true));
@@ -175,24 +219,35 @@ impl Todos {
         Ok(n)
     }
 
-    /// The agent marks one of its todos done, by id. Returns the number of
-    /// rows deleted (0 when the id was unknown / already gone).
+    /// The agent marks one of its todos done, by id — i.e. dismisses it from
+    /// its own list. Sets `acked` rather than deleting the row (see the
+    /// module doc comment for why): the row stays intact so a reconciled
+    /// producer's next `upsert` can still tell "unchanged" from "genuinely
+    /// different" instead of seeing an empty table and treating a re-poll of
+    /// the same condition as brand new. Returns the number of rows newly
+    /// acked (0 when the id was unknown, already gone, or already acked —
+    /// acking twice is not a new action).
     ///
     /// # Errors
     ///
-    /// Propagates the sqlite delete failure.
+    /// Propagates the sqlite update failure.
     ///
     /// # Panics
     ///
     /// Panics if the connection mutex is poisoned.
     pub fn mark_done(&self, id: i64) -> Result<usize> {
         let conn = self.conn.lock().unwrap();
-        let n = conn.execute("DELETE FROM todos WHERE id = ?1", params![id])?;
+        let n = conn.execute(
+            "UPDATE todos SET acked = 1, acked_at = ?1 WHERE id = ?2 AND acked = 0",
+            params![now_unix(), id],
+        )?;
         Ok(n)
     }
 
     /// List todos, newest-updated first. `subsystem = Some(..)` filters to
-    /// one producer's set; `None` returns all.
+    /// one producer's set; `None` returns all. Excludes acked rows — once
+    /// the agent has dismissed a todo it stays out of its own list, even
+    /// though the row itself survives (see `mark_done`).
     ///
     /// # Errors
     ///
@@ -206,7 +261,7 @@ impl Todos {
         let mut stmt = conn.prepare(
             "SELECT id, subsystem, subsystem_key, summary, source, updated_at \
              FROM todos \
-             WHERE (?1 IS NULL OR subsystem = ?1) \
+             WHERE acked = 0 AND (?1 IS NULL OR subsystem = ?1) \
              ORDER BY updated_at DESC, id DESC",
         )?;
         let rows = stmt
@@ -224,12 +279,12 @@ impl Todos {
         Ok(rows)
     }
 
-    /// Cheap existence check — `true` when at least one todo row exists
-    /// (across every subsystem). Used to gate a `todo_wake` notification
-    /// against being turned into a turn when its payload has already been
-    /// drained by an earlier turn (see the serve loop's `LocalTodo` arm):
-    /// an `EXISTS` probe, not a full `list` + row materialization, since
-    /// this runs on every wake.
+    /// Cheap existence check — `true` when at least one un-acked todo row
+    /// exists (across every subsystem). Used to gate a `todo_wake`
+    /// notification against being turned into a turn when its payload has
+    /// already been drained by an earlier turn (see the serve loop's
+    /// `LocalTodo` arm): an `EXISTS` probe, not a full `list` + row
+    /// materialization, since this runs on every wake.
     ///
     /// # Errors
     ///
@@ -240,10 +295,36 @@ impl Todos {
     /// Panics if the connection mutex is poisoned.
     pub fn has_any(&self) -> Result<bool> {
         let conn = self.conn.lock().unwrap();
-        let any: bool =
-            conn.query_row("SELECT EXISTS(SELECT 1 FROM todos)", [], |row| row.get(0))?;
+        let any: bool = conn.query_row(
+            "SELECT EXISTS(SELECT 1 FROM todos WHERE acked = 0)",
+            [],
+            |row| row.get(0),
+        )?;
         Ok(any)
     }
+
+    /// Reap acked rows whose `acked_at` is older than `cutoff` (a unix
+    /// timestamp). Called by the harness's hourly vacuum sweep
+    /// (`vacuum.rs`), not on any hot path — bounds the table against acked
+    /// one-shot todos (bash/matrix/forge) that will never naturally be
+    /// upserted again and so would otherwise sit forever. Never touches
+    /// un-acked rows. Returns the number of rows deleted.
+    ///
+    /// # Errors
+    ///
+    /// Propagates the sqlite delete failure.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the connection mutex is poisoned.
+    pub fn reap_acked(&self, cutoff: i64) -> Result<usize> {
+        let conn = self.conn.lock().unwrap();
+        let n = conn.execute(
+            "DELETE FROM todos WHERE acked = 1 AND acked_at < ?1",
+            params![cutoff],
+        )?;
+        Ok(n)
+    }
 }
 
 #[cfg(test)]
@@ -291,13 +372,20 @@ mod tests {
     }
 
     #[test]
-    fn clear_and_mark_done_remove_rows() {
+    fn clear_deletes_and_mark_done_hides_from_list() {
         let (_dir, s) = store();
         s.upsert("matrix", Some("!a:x"), "unread", None).unwrap();
         let (id, _) = s.upsert("forge", Some("pr-1"), "review", None).unwrap();
-        assert_eq!(s.clear("matrix", Some("!a:x")).unwrap(), 1);
-        assert_eq!(s.mark_done(id).unwrap(), 1);
-        assert!(s.list(None).unwrap().is_empty());
+        assert_eq!(s.clear("matrix", Some("!a:x")).unwrap(), 1, "clear deletes");
+        assert_eq!(
+            s.mark_done(id).unwrap(),
+            1,
+            "mark_done acks, doesn't delete"
+        );
+        assert!(
+            s.list(None).unwrap().is_empty(),
+            "both are gone from the agent-visible list either way"
+        );
     }
 
     #[test]
@@ -324,4 +412,101 @@ mod tests {
         assert_eq!(left.len(), 1);
         assert_eq!(left[0].subsystem, "forge");
     }
+
+    /// The whole point of the fix: acking a reconciled todo must not defeat
+    /// the producer's own anti-nag comparison. A re-push with an identical
+    /// summary after ack stays quiet AND stays out of the list — the `disk_watch`
+    /// bug this closes.
+    #[test]
+    fn acked_row_survives_unchanged_reupsert_and_stays_hidden() {
+        let (_dir, s) = store();
+        let (id, changed1) = s.upsert("disk", Some("usage"), "over 85%", None).unwrap();
+        assert!(changed1);
+        assert_eq!(s.mark_done(id).unwrap(), 1, "first ack succeeds");
+        assert!(s.list(None).unwrap().is_empty(), "acked row is hidden");
+
+        // Same producer, same reconcile tick, identical summary — must NOT
+        // look like a new todo just because the agent dismissed it.
+        let (id2, changed2) = s.upsert("disk", Some("usage"), "over 85%", None).unwrap();
+        assert_eq!(id, id2, "same row, not a fresh insert");
+        assert!(!changed2, "identical summary stays quiet even though acked");
+        assert!(
+            s.list(None).unwrap().is_empty(),
+            "still hidden — an unchanged re-push does not un-ack"
+        );
+    }
+
+    /// A materially different summary on an acked row is real information —
+    /// it must un-ack, re-surface in `list`, and report `changed = true`.
+    #[test]
+    fn acked_row_resurfaces_on_genuine_change() {
+        let (_dir, s) = store();
+        let (id, _) = s.upsert("disk", Some("usage"), "over 85%", None).unwrap();
+        s.mark_done(id).unwrap();
+        assert!(s.list(None).unwrap().is_empty());
+
+        let (id2, changed) = s.upsert("disk", Some("usage"), "over 90%", None).unwrap();
+        assert_eq!(id, id2);
+        assert!(changed, "a genuine re-deterioration must not be silenced");
+        let visible = s.list(None).unwrap();
+        assert_eq!(visible.len(), 1, "un-acked row is visible again");
+        assert_eq!(visible[0].summary, "over 90%");
+    }
+
+    /// Acking an already-acked (or unknown) id is a no-op, not a new action.
+    #[test]
+    fn mark_done_twice_is_idempotent() {
+        let (_dir, s) = store();
+        let (id, _) = s.upsert("bash", None, "task done", None).unwrap();
+        assert_eq!(s.mark_done(id).unwrap(), 1);
+        assert_eq!(s.mark_done(id).unwrap(), 0, "already acked");
+        assert_eq!(s.mark_done(999_999).unwrap(), 0, "unknown id");
+    }
+
+    /// `has_any` must also treat an acked-only table as empty — it gates a
+    /// wake against a payload the agent has already drained.
+    #[test]
+    fn has_any_excludes_acked_rows() {
+        let (_dir, s) = store();
+        let (id, _) = s.upsert("bash", None, "task done", None).unwrap();
+        assert!(s.has_any().unwrap());
+        s.mark_done(id).unwrap();
+        assert!(!s.has_any().unwrap(), "acked-only table reads as empty");
+    }
+
+    /// `reap_acked` only removes acked rows past the cutoff — a recent ack
+    /// and any un-acked row both survive.
+    #[test]
+    fn reap_acked_only_removes_old_acked_rows() {
+        let (dir, s) = store();
+        let (old_id, _) = s.upsert("bash", None, "old task", None).unwrap();
+        let (recent_id, _) = s.upsert("bash", None, "recent task", None).unwrap();
+        let (never_acked_id, _) = s.upsert("forge", Some("pr-1"), "review", None).unwrap();
+        s.mark_done(old_id).unwrap();
+        s.mark_done(recent_id).unwrap();
+
+        // Backdate the old row's ack past the cutoff directly — mark_done
+        // always acks "now", so simulate age via the raw connection.
+        {
+            let conn = s.conn.lock().unwrap();
+            conn.execute(
+                "UPDATE todos SET acked_at = ?1 WHERE id = ?2",
+                params![now_unix() - 1000, old_id],
+            )
+            .unwrap();
+        }
+
+        let removed = s.reap_acked(now_unix() - 500).unwrap();
+        assert_eq!(removed, 1, "only the backdated row is past the cutoff");
+
+        let conn = Connection::open(dir.path().join("todos.sqlite")).unwrap();
+        let remaining: i64 = conn
+            .query_row("SELECT COUNT(*) FROM todos", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(
+            remaining, 2,
+            "recent ack + never-acked row both survive the reap"
+        );
+        let _ = never_acked_id;
+    }
 }

hive-agent/src/vacuum.rs

@@ -32,6 +32,12 @@ const STREAM_KEEP_SECS: i64 = 14 * 24 * 3600;
 /// before reaping them — same window as `STREAM_KEEP_SECS`, kept around
 /// only to serve the trailing-window `ReminderRollup` stats.
 const REMINDER_KEEP_SECS: i64 = 14 * 24 * 3600;
+/// Keep acked (agent-dismissed) todo rows this long before reaping them —
+/// see `todos.rs`'s module doc for why acking doesn't delete outright. Long
+/// enough that a genuinely quiet month is the only way to trigger the "one
+/// spurious re-announcement" fallback path, short enough the table doesn't
+/// grow meaningfully from one-shot todos that will never be upserted again.
+const TODO_ACKED_KEEP_SECS: i64 = 30 * 24 * 3600;
 /// Terminal bash-task statuses whose files are eligible for deletion.
 const TERMINAL_STATUSES: &[&str] = &["done", "timed_out", "interrupted"];
 
@@ -72,6 +78,11 @@ fn sweep_once() {
             Ok(n) => tracing::info!(removed = n, "reminders vacuum"),
             Err(e) => tracing::warn!(error = ?e, "reminders vacuum failed"),
         }
+        match vacuum_todos(&state_db) {
+            Ok(0) => {}
+            Ok(n) => tracing::info!(removed = n, "todos vacuum"),
+            Err(e) => tracing::warn!(error = ?e, "todos vacuum failed"),
+        }
     }
 }
 
@@ -84,6 +95,13 @@ fn vacuum_reminders(path: &Path) -> anyhow::Result<usize> {
     store.prune_delivered_older_than(now_unix() - REMINDER_KEEP_SECS)
 }
 
+/// Reap acked todo rows older than [`TODO_ACKED_KEEP_SECS`] via the typed
+/// store API — same own-short-lived-connection shape as `vacuum_reminders`.
+fn vacuum_todos(path: &Path) -> anyhow::Result<usize> {
+    let store = crate::todos::Todos::open(path)?;
+    store.reap_acked(now_unix() - TODO_ACKED_KEEP_SECS)
+}
+
 /// Delete eligible bash-task trios in `dir`. Returns the count of `.json`
 /// sentinels removed (each represents one task; `.out`/`.err` deletions are
 /// not counted separately).