fleetd #284: free backend-error capacity

fleetd/src/main/java/dev/ltms/fleet/Fleetd.java

@@ -250,9 +250,7 @@ public final class Fleetd {
         boolean clearAfterTurn = cfg.lifecycle() != null && cfg.lifecycle().clearAfterTurn();
         SessionManager sessions = new SessionManager(workers, new GitWorktrees(cfg.worktreeRoot(), cfg.worktreeGroup()),
                 System::nanoTime, contextCap, clearAfterTurn);
-        liveCountRef.set(profileName -> (int) sessions.roster().stream()
-                .filter(s -> profileName.equals(s.profile()))
-                .count());
+        liveCountRef.set(profileName -> liveSessionCount(sessions.roster(), profileName));
 
         // CB-303 part 1: idle-ttl reaper — only when configured, defaults to disabled.
         final SessionReaper reaper;
@@ -859,6 +857,18 @@ public final class Fleetd {
                 .orElse(null);
     }
 
+    /**
+     * Count sessions that occupy a profile's spawn capacity. A {@code BACKEND_ERROR} session stays
+     * in the roster so {@code fleet_list} can show its failure, but its dead backend cannot use a
+     * seat or accept another delivery.
+     */
+    static int liveSessionCount(List<MemberSession> roster, String profileName) {
+        return (int) roster.stream()
+                .filter(session -> profileName.equals(session.profile()))
+                .filter(session -> session.state() != MemberSession.State.BACKEND_ERROR)
+                .count();
+    }
+
     /**
      * fleetd #248 / fleetd#201 Unit 5: factory for the production {@link BackendErrorSink} — the
      * collaborator {@link CompletionResolver} notifies when a pane-scrape classification actually

fleetd/src/main/java/dev/ltms/fleet/mcp/FleetMcp.java

@@ -1195,7 +1195,8 @@ public final class FleetMcp {
         int live = liveCount.apply(profile);
         int leadSeatCount = leadSeats.seatsFor().apply(profile);
         int reclaimable = (int) roster.stream().filter(s -> profile.equals(s.profile()))
-                .filter(s -> (s.state() == MemberSession.State.READY || s.state() == MemberSession.State.DONE))
+                .filter(s -> s.state() == MemberSession.State.READY || s.state() == MemberSession.State.DONE
+                        || s.state() == MemberSession.State.BACKEND_ERROR)
                 .filter(s -> messages == null || (!messages.hasAcceptedDelivery(s.terminalId()) && !messages.hasInboxMessage(s.terminalId())))
                 .count();
         Map<String, Object> row = new LinkedHashMap<>();

fleetd/src/test/java/dev/ltms/fleet/FleetdBackendErrorSinkTest.java

@@ -226,4 +226,27 @@ class FleetdBackendErrorSinkTest {
         assertTrue(remaining.isPresent(), "two distinct targets must start a cool-off");
         assertEquals(1, leadClient.sendCount());
     }
+
+    @Test
+    @DisplayName("a backend-error session no longer blocks the real maxLoad spawn gate")
+    void backendErrorSessionDoesNotBlockFreshSpawnAtMaxLoad() {
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig.Profile profile = new FleetConfig.Profile("terra", "http://gx00.gw:8000", "coder",
+                null, "FLEETD_WORKER_TOKEN", List.of("claude"), "tab", "fleetd-workers",
+                "w #{n}", null, null, null, null, null, null, null, 1.0f, 1);
+        Map<String, FleetConfig.Profile> profiles = Map.of("terra", profile);
+        ClaudeCodeLauncher adapter = new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), profiles, "terra", _ -> "tok");
+        AtomicReference<SessionManager> sessionsRef = new AtomicReference<>();
+        CompositePeerLauncher workers = new CompositePeerLauncher(List.of(adapter), "terra", profiles,
+                PlacementPolicies.fixed(), name -> Fleetd.liveSessionCount(sessionsRef.get().roster(), name));
+        SessionManager sessions = new SessionManager(workers);
+        sessionsRef.set(sessions);
+        MemberSession failed = sessions.acquire("terra", null, null, null);
+
+        assertTrue(sessions.onBackendError(failed.terminalId(), "backend exited"));
+
+        MemberSession fresh = sessions.acquire("terra", null, null, null);
+        assertEquals("terra", fresh.profile(), "the real maxLoad gate grants a fresh spawn after a backend error");
+    }
 }

fleetd/src/test/java/dev/ltms/fleet/mcp/FleetMcpTest.java

@@ -604,6 +604,21 @@ class FleetMcpTest {
         assertTrue(out.contains("\"reclaimable\":0"), out);
     }
 
+    @Test
+    void capacityReportsABackendErrorSessionAsReclaimable() {
+        FakeHerdr h = new FakeHerdr();
+        SessionManager sessions = new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")));
+        MemberSession session = sessions.acquire("ltms-local", null, null, null);
+        assertTrue(sessions.onBackendError(session.terminalId(), "backend exited"));
+
+        String out = textOf(FleetMcp.listFleet(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")),
+                sessions, null, new FleetMcp.CapacitySource(profile -> 0, profile -> 1,
+                () -> Set.of("ltms-local"), () -> 0), new FleetMcp.HealthCoverageSource(() -> "off"),
+                FleetMcp.QuarantineSource.none(), Map.of(), ""));
+
+        assertTrue(out.contains("\"reclaimable\":1"), out);
+    }
+
     @Test
     void inertCapacitySourceOmitsCapacityBlock() {
         FakeHerdr h = new FakeHerdr();