diff --git a/.agents/skills/fleet-dashboard/SKILL.md b/.agents/skills/fleet-dashboard/SKILL.md
index 6aac1d2e79a..7be49b0330b 100644
--- a/.agents/skills/fleet-dashboard/SKILL.md
+++ b/.agents/skills/fleet-dashboard/SKILL.md
@@ -35,10 +35,13 @@ Always pass `--prompt` (or `--prompt-file`) with his own words, unedited: that b
 `--captain` takes `firstmate`, `dj`, or `river` - whichever of the three is actually driving the work.
 New cards default to `not_started`; only set `--status` explicitly if work is already under way.
 
-## The six statuses
+## The seven statuses
 
-Use exactly these, never a synonym, so the board's filters and the auditor's checks stay meaningful:
+Use exactly these, never a synonym, so the board's filters and the auditor's checks stay meaningful.
 
+`needs-attention` and `testing` are easy to conflate and are opposite in the one way that matters: **`testing` is optional to him; `needs-attention` is blocking on him.** If he never looks at a `testing` card, nothing is lost - the work is otherwise complete. If he never looks at a `needs-attention` card, the work is stuck. Put a card in `needs-attention` only when it genuinely cannot proceed without him - a decision, an answer, a physical action, something only he can do or judge. A card that is simply done and could use his eyes stays in `testing`.
+
+- `needs-attention` - the work cannot proceed without him: a decision, an answer, a physical action, anything only he can supply. Use `--reason` to say what, briefly - that text renders directly on the card so he can act without opening it. This is the loudest status on the board and sorts first; do not use it for routine progress or for something an agent could resolve on its own.
 - `not-started` - received, nothing begun yet.
 - `working` - an agent is actively on it right now, not "queued" or "about to start."
 - `paused` - the Admiral paused it. Only he pauses a task; do not set this because a crew went idle for another reason - that is a stalled task, not a paused one, and belongs in a status update or the auditor's discrepancy log instead.
@@ -73,11 +76,12 @@ If you are the fleet auditor (or standing in for it), your job every cycle is to
 1. `bin/fm-dashboard.sh list --status working --json` - for each, confirm a real agent is actually on it (live crew/session state, not the card's own say-so). Anything claiming `working` with no corroborating live activity is a genuine discrepancy.
 2. `bin/fm-dashboard.sh list --status waiting --json` - confirm the blocker is still real. If a `waiting_on_id` card is already `complete`, or the external condition it names has cleared, that is a discrepancy: the card is stale, not honestly waiting.
 3. `bin/fm-dashboard.sh list --status paused --json` - confirm each is still genuinely paused by the Admiral's own word, not just quiet. If one has been unpaused, confirm it is *actually being worked* - an unpaused-but-idle task is exactly the failure this check exists to catch.
-4. For anything you log as a discrepancy, use `bin/fm-dashboard.sh audit-log <id> "<what you found>"`. Be concrete: state what the card claims and what you actually observed, so firstmate can act without re-deriving it.
-5. If you cannot verify a card at all (no linked reference, no way to check from where you sit), that is **not** a discrepancy - do not log it as one. Silence on an unverifiable card is correct; a false "wrong" trains the Admiral to ignore the log exactly the way an always-red marker already has once.
-6. When the sweep finishes, always call `bin/fm-dashboard.sh audit-run --duration-seconds <n> --checked <n> --discrepancies <n>`, even when nothing was wrong. This is what lets the page show "clean" as a real, timed answer instead of an absence.
-7. If the sweep itself fails partway (a source you needed was unreadable, a check errored out), log that with `--kind error` via `audit-log --fleet "<what broke>"` rather than silently posting a shorter, quieter run. A failed check must never look identical to a clean one.
-8. Read the current cadence with `bin/fm-dashboard.sh audit-interval get` at the start of each cycle rather than assuming the last-known value; the Admiral can change it from the page at any time.
+4. `bin/fm-dashboard.sh list --status needs-attention --json` - check each `needs_attention` card's own `show <id>` for how long it has actually sat in that status (its status history has the timestamp it last changed to `needs_attention`). **Age itself is the finding here, asymmetrically with every other status**: a `needs-attention` card that has sat for hours with no reply from him means he was not asked clearly, or the ask never reached him - log it even when the card's claim is otherwise accurate. A `testing` card sitting for the same length of time is not a discrepancy at all - `testing` is optional to him by design, so its age proves nothing. Do not apply this age check to any status but `needs-attention`.
+5. For anything you log as a discrepancy, use `bin/fm-dashboard.sh audit-log <id> "<what you found>"`. Be concrete: state what the card claims and what you actually observed, so firstmate can act without re-deriving it.
+6. If you cannot verify a card at all (no linked reference, no way to check from where you sit), that is **not** a discrepancy - do not log it as one. Silence on an unverifiable card is correct; a false "wrong" trains the Admiral to ignore the log exactly the way an always-red marker already has once.
+7. When the sweep finishes, always call `bin/fm-dashboard.sh audit-run --duration-seconds <n> --checked <n> --discrepancies <n>`, even when nothing was wrong. This is what lets the page show "clean" as a real, timed answer instead of an absence.
+8. If the sweep itself fails partway (a source you needed was unreadable, a check errored out), log that with `--kind error` via `audit-log --fleet "<what broke>"` rather than silently posting a shorter, quieter run. A failed check must never look identical to a clean one.
+9. Read the current cadence with `bin/fm-dashboard.sh audit-interval get` at the start of each cycle rather than assuming the last-known value; the Admiral can change it from the page at any time.
 
 ## What not to do
 
diff --git a/bin/fleet-dashboard/server/store.py b/bin/fleet-dashboard/server/store.py
index 1b130e8ec3c..4434fc9b4c6 100644
--- a/bin/fleet-dashboard/server/store.py
+++ b/bin/fleet-dashboard/server/store.py
@@ -17,7 +17,7 @@
 import time
 from contextlib import contextmanager
 
-STATUSES = ("not_started", "working", "paused", "waiting", "testing", "complete")
+STATUSES = ("needs_attention", "not_started", "working", "paused", "waiting", "testing", "complete")
 CAPTAINS = ("firstmate", "captain_dj", "captain_river")
 NOTE_TABS = ("interpretation", "communication", "needs")
 NOTE_AUTHORS = ("agent", "firstmate", "admiral")
@@ -31,6 +31,7 @@
   status TEXT NOT NULL,
   waiting_on_id TEXT,
   waiting_reason TEXT,
+  needs_attention_reason TEXT,
   starred INTEGER NOT NULL DEFAULT 0,
   backlog_ref TEXT,
   initial_prompt TEXT NOT NULL,
@@ -85,6 +86,14 @@
 
 DEFAULT_AUDIT_INTERVAL_MINUTES = 15
 
+# Columns added after the initial schema. CREATE TABLE IF NOT EXISTS leaves an
+# already-created table untouched, so a database created before a column
+# existed needs it added explicitly - this keeps existing cards and their
+# history intact instead of requiring a fresh database.
+_TASK_COLUMN_MIGRATIONS = (
+    ("needs_attention_reason", "TEXT"),
+)
+
 _write_lock = threading.Lock()
 
 
@@ -100,6 +109,10 @@ def __init__(self, db_path: str):
             os.makedirs(directory, exist_ok=True)
         with self._connect() as conn:
             conn.executescript(SCHEMA)
+            existing_columns = {row[1] for row in conn.execute("PRAGMA table_info(tasks)")}
+            for name, coltype in _TASK_COLUMN_MIGRATIONS:
+                if name not in existing_columns:
+                    conn.execute(f"ALTER TABLE tasks ADD COLUMN {name} {coltype}")
             conn.execute(
                 "INSERT OR IGNORE INTO settings(key, value) VALUES ('audit_interval_minutes', ?)",
                 (str(DEFAULT_AUDIT_INTERVAL_MINUTES),),
@@ -248,15 +261,19 @@ def set_status(self, task_id: str, status: str, waiting_on_id: str | None = None
             raise KeyError(task_id)
         if waiting_on_id and not self.task_exists(waiting_on_id):
             raise ValueError(f"waiting_on_id does not exist: {waiting_on_id!r}")
+        # `reason` is repurposed per status: what a card is waiting on for
+        # `waiting`, what is being asked of him for `needs_attention`. The two
+        # are mutually exclusive, so only the active status's column is kept.
+        waiting_reason = reason if status == "waiting" else None
+        needs_attention_reason = reason if status == "needs_attention" else None
         if status != "waiting":
             waiting_on_id = None
-            reason = reason if reason else None
         ts = now_iso()
         with self._cursor(write=True) as cur:
             cur.execute(
                 """UPDATE tasks SET status = ?, waiting_on_id = ?, waiting_reason = ?,
-                   updated_at = ? WHERE id = ?""",
-                (status, waiting_on_id, reason, ts, task_id),
+                   needs_attention_reason = ?, updated_at = ? WHERE id = ?""",
+                (status, waiting_on_id, waiting_reason, needs_attention_reason, ts, task_id),
             )
             cur.execute(
                 """INSERT INTO status_history (task_id, from_status, to_status, changed_at, note)
diff --git a/bin/fleet-dashboard/web/app.js b/bin/fleet-dashboard/web/app.js
index 57817915d73..f0c8790af38 100644
--- a/bin/fleet-dashboard/web/app.js
+++ b/bin/fleet-dashboard/web/app.js
@@ -11,8 +11,9 @@ import htm from "./vendor/htm.module.js";
 const { useState, useEffect, useCallback, useMemo, useRef } = React;
 const html = htm.bind(React.createElement);
 
-const STATUSES = ["not_started", "working", "paused", "waiting", "testing", "complete"];
+const STATUSES = ["needs_attention", "not_started", "working", "paused", "waiting", "testing", "complete"];
 const STATUS_META = {
+  needs_attention: { label: "Needs Attention" },
   not_started: { label: "Not Started" },
   working: { label: "Working" },
   paused: { label: "Paused" },
@@ -121,6 +122,11 @@ function Card({ task, allTasks, onOpen, onToggleStar, onQuickStatus }) {
         <${StatusPill} status=${task.status} />
         <${CaptainPill} captain=${task.captain} />
       </div>
+      ${task.status === "needs_attention" ? html`
+        <div class="needs-attention-banner">
+          ⚑ ${task.needs_attention_reason || "Needs his decision or action to continue."}
+        </div>
+      ` : null}
       ${task.status === "waiting" ? html`
         <div>
           ${waitingOn
@@ -574,8 +580,12 @@ function App() {
     return sorted;
   }, [tasks, filters]);
 
-  const favorites = useMemo(() => filtered.filter((t) => t.starred), [filtered]);
-  const rest = useMemo(() => filtered.filter((t) => !t.starred), [filtered]);
+  // Needs Attention is pulled out of every other section - blocking-on-him
+  // work must be the first thing he sees, never buried among starred or
+  // routine cards regardless of the active sort or filter.
+  const needsAttention = useMemo(() => filtered.filter((t) => t.status === "needs_attention"), [filtered]);
+  const favorites = useMemo(() => filtered.filter((t) => t.starred && t.status !== "needs_attention"), [filtered]);
+  const rest = useMemo(() => filtered.filter((t) => !t.starred && t.status !== "needs_attention"), [filtered]);
 
   const counts = useMemo(() => {
     const status = {}, captain = {};
@@ -598,6 +608,19 @@ function App() {
 
       <${ConnBanner} error=${connError} lastOkAt=${lastOkAt} />
 
+      ${needsAttention.length > 0 ? html`
+        <h2 class="section needs-attention-section">Needs Attention <span class="count">(${needsAttention.length})</span></h2>
+        <div class="grid needs-attention-grid">
+          ${needsAttention.map((t) => html`
+            <${Card}
+              key=${t.id} task=${t} allTasks=${tasks}
+              onOpen=${setSelectedId} onToggleStar=${toggleStar}
+              onQuickStatus=${(task, status) => setStatus(task.id, status)}
+            />
+          `)}
+        </div>
+      ` : null}
+
       ${favorites.length > 0 ? html`
         <h2 class="section">Favorites <span class="count">(${favorites.length})</span></h2>
         <div class="favorites-row">
diff --git a/bin/fleet-dashboard/web/styles.css b/bin/fleet-dashboard/web/styles.css
index 6c44133528c..4036bfc2992 100644
--- a/bin/fleet-dashboard/web/styles.css
+++ b/bin/fleet-dashboard/web/styles.css
@@ -12,12 +12,14 @@
   --font: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Inter, system-ui, sans-serif;
 
   /* status colors */
+  --st-needs_attention: #ff3b30;
   --st-not_started: #7a8699;
   --st-working: #1db954;
   --st-paused: #e0a940;
   --st-waiting: #a970ff;
   --st-testing: #ffd700;
   --st-complete: #5b8cff;
+  --st-needs_attention-bg: rgba(255, 59, 48, .12);
 
   /* captain tag colors */
   --cap-firstmate: #9a9aa2;
@@ -139,6 +141,10 @@ h2.section .count { color: var(--text); font-weight: 700; }
   display: flex; flex-direction: column; gap: 8px; min-width: 0;
 }
 .card:hover { border-color: #4a4a4a; }
+.card.st-needs_attention {
+  border-left-width: 6px; border-left-color: var(--st-needs_attention);
+  border-color: var(--st-needs_attention); background: var(--st-needs_attention-bg);
+}
 .card.st-not_started { border-left-color: var(--st-not_started); }
 .card.st-working { border-left-color: var(--st-working); }
 .card.st-paused { border-left-color: var(--st-paused); }
@@ -160,6 +166,7 @@ h2.section .count { color: var(--text); font-weight: 700; }
   padding: 2.5px 8px; border-radius: 99px; white-space: nowrap;
 }
 .pill.status { color: #06110b; }
+.pill.status.st-needs_attention { background: var(--st-needs_attention); color: #2b0503; }
 .pill.status.st-not_started { background: var(--st-not_started); color: #0c1116; }
 .pill.status.st-working { background: var(--st-working); }
 .pill.status.st-paused { background: var(--st-paused); color: #221a04; }
@@ -175,6 +182,16 @@ h2.section .count { color: var(--text); font-weight: 700; }
 .card-meta { font-size: 12px; color: var(--text-muted); display: flex; justify-content: space-between; gap: 8px; }
 .card-agent { color: var(--text); font-weight: 500; }
 
+.needs-attention-banner {
+  font-size: 13px; font-weight: 700; color: #ffb4ac;
+  background: var(--danger-bg); border: 1px solid var(--st-needs_attention);
+  border-radius: 6px; padding: 8px 10px;
+}
+
+h2.section.needs-attention-section { color: var(--st-needs_attention); }
+h2.section.needs-attention-section::after { background: var(--st-needs_attention); opacity: .4; }
+.needs-attention-grid { margin-top: 10px; margin-bottom: 6px; }
+
 .waiting-link-btn, .quick-action {
   font-size: 12px; font-weight: 600; border-radius: 6px; padding: 6px 10px;
   border: 1px solid var(--border); background: var(--surface2); color: var(--text); cursor: pointer;
diff --git a/bin/fm-dashboard.sh b/bin/fm-dashboard.sh
index 0c146a0988a..3ea957d0399 100755
--- a/bin/fm-dashboard.sh
+++ b/bin/fm-dashboard.sh
@@ -24,6 +24,8 @@
 #   fm-dashboard.sh captain <id> <firstmate|dj|river>
 #   fm-dashboard.sh ref <id> <backlog-ref>
 #   fm-dashboard.sh status <id> <status> [--waiting-on <id>] [--reason <text>]
+#       --reason is what the card is waiting on for `waiting`, or what is
+#       being asked of him for `needs-attention`; ignored for other statuses.
 #   fm-dashboard.sh star <id>
 #   fm-dashboard.sh unstar <id>
 #   fm-dashboard.sh note <id> --tab <interpretation|communication|needs> \
@@ -36,7 +38,7 @@
 #   fm-dashboard.sh start|stop|restart|server-status   (server process lifecycle)
 #   fm-dashboard.sh --help
 #
-# statuses: not-started working paused waiting testing complete
+# statuses: needs-attention not-started working paused waiting testing complete
 # tabs:     interpretation communication needs
 # captain shorthands: dj -> captain_dj, river -> captain_river, firstmate -> firstmate
 #
@@ -110,8 +112,9 @@ json_escape() { need_tool jq; jq -Rs . <<<"$1"; }
 canon_status() {
   case "$1" in
     not-started|not_started) printf 'not_started' ;;
+    needs-attention|needs_attention) printf 'needs_attention' ;;
     working|paused|waiting|testing|complete) printf '%s' "$1" ;;
-    *) die "unknown status '$1' - valid: not-started working paused waiting testing complete" ;;
+    *) die "unknown status '$1' - valid: needs-attention not-started working paused waiting testing complete" ;;
   esac
 }
 
@@ -203,6 +206,7 @@ cmd_show() {
     "agent:    \(.agent)",
     "starred:  \(.starred == 1)",
     (if .status == "waiting" then "waiting on: \(.waiting_on_id // "(no card)") - \(.waiting_reason // "")" else empty end),
+    (if .status == "needs_attention" then "needs attention: \(.needs_attention_reason // "(no reason recorded)")" else empty end),
     (if .backlog_ref then "ref:      \(.backlog_ref)" else empty end),
     "",
     "--- prompt ---",
@@ -432,7 +436,7 @@ main() {
     stop) cmd_server_stop ;;
     restart) cmd_server_stop 2>/dev/null; cmd_server_start ;;
     server-status) cmd_server_status ;;
-    ""|--help|-h|help) sed -n '2,45p' "${BASH_SOURCE[0]}" | sed 's/^# \{0,1\}//' ;;
+    ""|--help|-h|help) sed -n '2,48p' "${BASH_SOURCE[0]}" | sed 's/^# \{0,1\}//' ;;
     *) die "unknown command '$cmd' - run: fm-dashboard.sh --help" ;;
   esac
 }
diff --git a/docs/dashboard.md b/docs/dashboard.md
index 8cb3380c35a..deb7caed83b 100644
--- a/docs/dashboard.md
+++ b/docs/dashboard.md
@@ -1,7 +1,7 @@
 # The Admiral's Fleet Dashboard
 
 A purpose-built task board, styled to match Spectra, that replaced the generated Lavish status page (`bin/fm-status-board.sh`) as the Admiral's primary fleet surface.
-One card per task the fleet has been given; six statuses; a captain tag for who is driving it; four tabs per card; a bottom-of-page discrepancy log kept by the fleet auditor.
+One card per task the fleet has been given; seven statuses; a captain tag for who is driving it; four tabs per card; a bottom-of-page discrepancy log kept by the fleet auditor.
 
 For exact current command syntax, run `bin/fm-dashboard.sh --help`.
 For when and why an agent calls which command, see [`fleet-dashboard`](../.agents/skills/fleet-dashboard/SKILL.md) - that skill is the single owner of agent-facing usage guidance; this page stays architecture, setup, and the decisions behind the shape.
@@ -56,6 +56,14 @@ This was a deliberate choice among three options - read live from the backlog, o
 The fleet auditor exists specifically to bound that risk: every sweep re-derives ground truth from live crew/session state and the real backlog, compares it to what each card claims, and logs a discrepancy the moment the two disagree, with a timed record of how long the check took so a silently-skipped sweep is itself visible (see "Auditor integration" below).
 An optional `backlog_ref` field on a card (`bin/fm-dashboard.sh ref <id> <home:task-id>`) lets the auditor cross-check a specific card against a specific backlog entry when one exists; a card with no ref is not treated as wrong for that - see the next section.
 
+## Why `needs-attention` is a separate status from `testing`
+
+Both statuses put a finished-enough card in front of the Admiral, which is why they used to get conflated - and why doing so once buried several of his genuinely open decisions in a place he had no reason to check closely.
+The two are opposite on the one axis that matters: whether the work still needs him to move forward.
+`testing` is done and optional - if he never opens the card, nothing is lost, it is otherwise complete.
+`needs-attention` is stuck without him - a decision, an answer, or a physical action only he can supply.
+That asymmetry is why `needs-attention` sorts first and renders loudest on the page, and why the fleet auditor treats its age as a finding in a way it never does for `testing` (see the [`fleet-dashboard`](../.agents/skills/fleet-dashboard/SKILL.md) skill for the exact status definitions and the auditor's per-status procedure).
+
 ## Link policy (standing order 17)
 
 `bin/fm-dashboard.sh link` and the underlying `POST /api/tasks/{id}/notes` endpoint reject, structurally, any link whose host contains `github`, any link that is not a full `http(s)://` URL, and any link whose host is local-only and will not resolve from the Admiral's phone (`bin/fleet-dashboard/server/validation.py`).
diff --git a/tests/fm-dashboard.test.sh b/tests/fm-dashboard.test.sh
index e1bfc3c9db4..19fbfd87029 100755
--- a/tests/fm-dashboard.test.sh
+++ b/tests/fm-dashboard.test.sh
@@ -94,6 +94,27 @@ test_waiting_status_carries_target_and_reason() {
   pass "waiting status carries its target card and reason"
 }
 
+test_needs_attention_status_carries_reason_and_sorts_first() {
+  local id working_id out
+  id=$("$DASH" add --title "Needs a decision" --captain firstmate --prompt "checking needs-attention" | awk '{print $1}')
+  working_id=$("$DASH" add --title "Being actively worked" --captain firstmate --prompt "sort-order control" --status working | awk '{print $1}')
+
+  "$DASH" status "$id" needs-attention --reason "pick red or blue for the trim" >/dev/null \
+    || fail "status transition to needs-attention failed"
+  out=$("$DASH" show "$id")
+  assert_contains "$out" "status:   needs_attention" "needs-attention status did not persist"
+  assert_contains "$out" "needs attention: pick red or blue for the trim" "needs-attention reason did not persist"
+
+  local first_id
+  first_id=$("$DASH" list --sort status | head -n1 | awk '{print $1}')
+  [ "$first_id" = "$id" ] || fail "needs-attention ($id) did not sort above a working card ($working_id) under --sort status, got: $first_id"
+
+  "$DASH" status "$id" working >/dev/null || fail "leaving needs-attention failed"
+  assert_not_contains "$("$DASH" show "$id")" "needs attention:" "needs-attention reason was not cleared on status change"
+
+  pass "needs-attention status carries a reason and sorts above every other status"
+}
+
 test_star_and_delete() {
   local id
   id=$(cat "$FM_HOME/task-id")
@@ -181,6 +202,7 @@ test_status_and_captain_and_title_updates
 test_waiting_status_carries_target_and_reason
 test_notes_tabs_and_empty_tab_semantics
 test_link_policy_rejects_github_and_localhost
+test_needs_attention_status_carries_reason_and_sorts_first
 test_audit_log_run_and_interval
 test_bad_input_fails_with_nonzero_exit
 test_star_and_delete