fleetd #302: require content in MessageService.reply so a REST reply with no content cannot silently resolve a waiter

release() cancelled the consumer and dropped its local record of deliveries the
broker still held as outstanding. Cancelling a consumer does not requeue them:
they stay unacked on the still-open channel until it or the connection closes.
So a held-but-undrained reply became permanently unreachable — a worker's
report lost with no error and no log line. It now nacks with requeue, after
cancelling, so a later own() can still receive it.

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

@@ -625,6 +625,19 @@ public final class Fleetd {
             log.info("auth: loopback-trust (any loopback non-worker caller is the primary)");
         }
 
+        // fleetd #297: named once and reused verbatim below for FleetApp's GET /profiles, rather than
+        // built a second time — two independently-constructed sources reading the SAME BackendQuarantine
+        // / BackendOutagePolicy would still be able to drift (e.g. a future edit to the credentialIdFor
+        // closure in only one of the two places), exactly the shape #284 was.
+        FleetMcp.QuarantineSource quarantineSource = new FleetMcp.QuarantineSource(profile -> {
+            var configured = config.get().profiles().get(profile);
+            return configured == null ? null : configured.effectiveCredentialId();
+        }, quarantine);
+        FleetMcp.OutageSource outageSource = new FleetMcp.OutageSource(profile -> {
+            var configured = config.get().profiles().get(profile);
+            return configured == null ? null : configured.effectiveCredentialId();
+        }, outagePolicy);
+
         FleetMcp mcp = new FleetMcp(messages, workers, sessions, identity, presence,
                 primaryRegistry, callers, metrics, new FleetMcp.CapacitySource(profile -> liveCountRef.get().apply(profile),
                 profile -> {
@@ -636,15 +649,9 @@ public final class Fleetd {
                     return FleetHealthMonitor.coverage(health != null && health.isEnabled(),
                             health != null && health.notifications() != null && health.notifications().configured());
                 }),
-                new FleetMcp.QuarantineSource(profile -> {
-                    var configured = config.get().profiles().get(profile);
-                    return configured == null ? null : configured.effectiveCredentialId();
-                }, quarantine),
+                quarantineSource,
                 leadMailbox,
-                new FleetMcp.OutageSource(profile -> {
-                    var configured = config.get().profiles().get(profile);
-                    return configured == null ? null : configured.effectiveCredentialId();
-                }, outagePolicy),
+                outageSource,
                 new FleetMcp.LeadSeatSource(leadSeatLookup(() -> config.get().profiles(), leaders, leads)));
 
         // CB-637: the receive half. Only constructed when a lead mailbox actually opened — with no
@@ -715,9 +722,12 @@ public final class Fleetd {
         // GET /sessions must merge across both, or a down/unpolled member daemon is invisible.
         // fleetd #111: live (re-read-per-request) memberCredentials view for GET /member-credentials —
         // same hot-reload shape as the memberCredentials supplier passed to ClaudeCodeLauncher above.
+        // fleetd #297: quarantineSource/outageSource are the SAME instances passed to FleetMcp above —
+        // GET /profiles must report the identical quarantine/cool-off facts as fleet_profiles.
         Javalin app = new FleetApp(herdr, memberHerdr, workers, sessions, messages, presence, mcp.servlet(),
                 callers, metrics, deliverable,
-                () -> MemberCredentialPolicyView.of(config.get().memberCredentials())).build();
+                () -> MemberCredentialPolicyView.of(config.get().memberCredentials()),
+                quarantineSource, outageSource).build();
         app.start(cfg.bind().host(), cfg.bind().port());
         log.info("fleetd listening on {}:{}, herdr socket {}",
                 cfg.bind().host(), cfg.bind().port(), socket);

fleetd/src/main/java/dev/ltms/fleet/health/FleetHealthMonitor.java

@@ -14,6 +14,7 @@ import java.util.HashSet;
 import java.util.List;
 import java.util.Map;
 import java.util.Objects;
+import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.ScheduledExecutorService;
 import java.util.concurrent.TimeUnit;
 import java.util.function.BiConsumer;
@@ -46,7 +47,16 @@ public final class FleetHealthMonitor {
     private final long workingSuspectAfterNanos;
     private final BiConsumer<String, String> failTarget;
     private final Map<String, HealthPrior> priors = new HashMap<>();
-    private final Map<String, HealthState> states = new HashMap<>();
+    /**
+     * The live classification per member, and the only one of this class's three maps that more
+     * than one scheduler task touches. {@code tick} writes it (and prunes it to the roster);
+     * fleetd #280's delayed {@link #recheckTerminalTarget} reads it from its own separate scheduled
+     * task. Both run on the single-threaded scheduler {@code Fleetd} passes in today, so they are
+     * serialised — but nothing in this class enforces that, and an unsynchronised {@link HashMap}
+     * read racing a resize can spin a CPU forever rather than fail visibly. {@code priors} and
+     * {@code orphanStreaks} stay plain maps because {@code tick} is still their only toucher.
+     */
+    private final Map<String, HealthState> states = new ConcurrentHashMap<>();
     /**
      * CB-643: consecutive ticks on which a target looked like an orphaned delegation. The fact
      * {@link MessageService#hasOrphanedDelegation} reports is a true snapshot, but it can read true

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

@@ -1015,6 +1015,25 @@ public final class FleetMcp {
      * both maps at once when it is both exhaustion-quarantined AND cooling off.
      */
     static McpSchema.CallToolResult profiles(PeerLauncher workers, QuarantineSource quarantine, OutageSource outage) {
+        return text(json(profilesView(workers, quarantine, outage)));
+    }
+
+    /**
+     * The body both front doors answer {@code profiles} with: the configured profile names, the
+     * default, and the two independent outage states — {@code quarantined} (the backend reported it
+     * out of capacity) and {@code coolingOff} (the credential threw repeated non-exhaustion backend
+     * errors). Each map is present only when at least one profile is in that state, and a profile
+     * can appear in both at once, because the two checks are separate.
+     *
+     * <p>fleetd #297: extracted so {@code fleet_profiles} and {@code GET /profiles} render from ONE
+     * body builder rather than two copies. Passing both doors the same {@link QuarantineSource} and
+     * {@link OutageSource} instances is necessary but not sufficient: with the loop written out
+     * twice, a later edit to the row shape — a renamed key, an added field — lands on one door and
+     * not the other, and the two then disagree about a live outage. That is exactly what fleetd
+     * #284 was, where one rule computed in two places was widened in only one and a single response
+     * contradicted itself. Shared inputs do not make duplicated computation safe.
+     */
+    public static Map<String, Object> profilesView(PeerLauncher workers, QuarantineSource quarantine, OutageSource outage) {
         Map<String, Object> result = new LinkedHashMap<>();
         result.put("profiles", workers.profiles());
         result.put("default", workers.defaultProfile() == null ? "" : workers.defaultProfile());
@@ -1046,7 +1065,7 @@ public final class FleetMcp {
         if (!coolingOff.isEmpty()) {
             result.put("coolingOff", coolingOff);
         }
-        return text(json(result));
+        return result;
     }
 
     /**

fleetd/src/main/java/dev/ltms/fleet/member/HerdrPeerLauncher.java

@@ -697,7 +697,20 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         if (paneId == null) {
             throw new IllegalStateException("pane.split returned no pane — cannot start a peer");
         }
-        Agent peer = startUniquelyNamed(cfg, argv, paneId).agent();
+        Agent peer;
+        try {
+            peer = startUniquelyNamed(cfg, argv, paneId).agent();
+        } catch (RuntimeException e) {
+            // The peer never started — don't leave the pane we just created orphaned.
+            // Best-effort cleanup; never let it mask the real spawn failure.
+            try {
+                stop(paneId);
+            } catch (RuntimeException cleanup) {
+                log.warn("failed to close orphaned pane {} after spawn error: {}",
+                        paneId, cleanup.getMessage());
+            }
+            throw e;
+        }
         log.info("{} started pane={} terminal={}", namePrefix, peer.paneId(), peer.terminalId());
         return peer;
     }
@@ -913,9 +926,14 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * {@link #reapOrphanWorkers() orphan-reap} and spawn-gate-timeout paths, plus any caller that
      * passes a pane directly, keep working without an owning id.
      *
-     * <p>Resolves the tab from the pane <em>before</em> closing it. An already-gone pane/tab
-     * (repeated DELETE, crashed peer) is treated as success; any other failure propagates so a
-     * genuinely failed teardown is not reported as done.
+     * <p>Resolves the tab from the pane <em>before</em> closing it. {@code agents.close} (the pane)
+     * is the one step whose failure means the teardown itself may not have happened: an already-gone
+     * pane (repeated DELETE, crashed peer) is treated as success, but any other failure propagates so
+     * a genuinely failed teardown is not reported as done. {@code spaces.closeTab} (fleetd #293) is
+     * different — by the time it runs the pane is already closed, so it is cosmetic workspace tidying
+     * rather than a real teardown failure, and a failure there is logged and never propagates, so it
+     * cannot mask the two cleanups below it ({@link #releaseZdotdir}, and the caller's worktree
+     * removal in {@code SessionManager.release}).
      */
     @Override
     public void stop(String idOrPane) {
@@ -934,7 +952,25 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
             log.debug("pane.close({}) ignored — already gone: {}", paneId, e.getMessage());
         }
         if (loc != null && loc.tabPaneCount() == 1) {
-            spaces.closeTab(loc.tabId());
+            // fleetd #293: the pane above is already closed by this point, so a failing tab.close is
+            // cosmetic workspace tidying, not a real teardown failure — it must not mask the two
+            // cleanups below it (releaseZdotdir, and the caller's worktree removal). Unlike
+            // agents.close above, this is not narrowed to "already gone": any failure here, whatever
+            // its cause, is one we continue past, so we log it at WARN (not debug) with the tab id a
+            // person can go close by hand.
+            try {
+                spaces.closeTab(loc.tabId());
+            } catch (RuntimeException e) {
+                // Caught as RuntimeException, not HerdrException, to match releaseZdotdir's own
+                // guard five lines below. Today the two are the same set — HerdrCodec wraps every
+                // encode/decode failure and UnixSocketHerdrClient wraps every IOException, so
+                // HerdrException is all closeTab can actually throw. Narrowing to it anyway would
+                // leave this step guarded against the expected failure and bare against any other,
+                // which is the exact asymmetry fleetd #293 exists to remove. No behaviour change
+                // today; it stops a later change inside WorkspaceControl.closeTab reopening it.
+                log.warn("tab.close({}) failed — the pane is already torn down, so continuing; the "
+                        + "tab may need manual cleanup: {}", loc.tabId(), e.getMessage());
+            }
         } else if (loc != null) {
             log.debug("not closing tab {} — it holds {} panes (not a dedicated peer tab)",
                     loc.tabId(), loc.tabPaneCount());
@@ -973,7 +1009,8 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * that gap: it stops waiting immediately (never burns the rest of the timeout), runs the same
      * teardown the timeout path below runs, and throws with a message that says the backend exited
      * rather than that the pane was slow. Any other {@link HerdrException} still propagates
-     * unchanged — this gate does not know how to recover from it.
+     * unchanged — this gate does not interpret or recover from it, but it still closes the pane
+     * it opened before handing the exception to its caller.
      */
     private void waitUntilInjectableOrThrow(String paneId) {
         long start = nowMillis.getAsLong();
@@ -987,7 +1024,15 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
                 if (isAlreadyGone(e)) {
                     failFastOnGoneBackend(paneId, e, nowMillis.getAsLong() - start);
                 }
-                throw e; // any other herdr failure is not ours to interpret — let it propagate
+                // This gate must not interpret an unrelated herdr error, but the caller does not
+                // receive paneId when spawn throws. Close the pane here before propagating e unchanged.
+                try {
+                    stop(paneId);
+                } catch (RuntimeException cleanup) {
+                    log.warn("failed to close orphaned pane {} after readiness-gate error: {}",
+                            paneId, cleanup.getMessage());
+                }
+                throw e;
             }
             lastStatus = sample.status();
             if (lastStatus.injectable() || refinedInjectable(paneId, sample)) {

fleetd/src/main/java/dev/ltms/fleet/msg/AmqpReplyInbox.java

@@ -208,18 +208,72 @@ public final class AmqpReplyInbox implements ReplyInbox, AutoCloseable {
         }
     }
 
+    /**
+     * Release ownership of {@code target}: cancel its consumer, then nack-with-requeue every
+     * delivery still held for it instead of just dropping the local record.
+     *
+     * <p><strong>Cancelling a consumer does not requeue its in-flight deliveries.</strong> In AMQP,
+     * a delivery that was pushed to a consumer stays unacked, attached to the still-open
+     * {@link #channel}, until that channel or the connection closes — {@code basicCancel} alone does
+     * neither. So before this method existed with a requeue step, it dropped {@link #held}'s entries
+     * for {@code target} while the broker still considered them outstanding: never acked, never
+     * nacked, never requeued, and no longer reachable by {@link #peek} — permanently invisible. This
+     * is unlike {@link #handleRecovery} and {@link #close()}, whose bare {@code held.clear()} is
+     * correct because each has already made the broker requeue (a real connection drop, or
+     * {@code channel.close()} respectively) before clearing local state.
+     *
+     * <p><strong>Order: cancel first, then nack.</strong> A delivery tag stays valid for
+     * {@code basicNack} on this channel regardless of whether its consumer is still attached — only
+     * a channel/connection close invalidates it — so cancelling {@code target}'s consumer first does
+     * not risk the tags. Doing it the other way round does: nacking a delivery with {@code requeue}
+     * while its consumer is still active hands the message straight back to that <em>same</em>
+     * consumer the instant a prefetch slot frees up (confirmed against a real broker — see
+     * {@code AmqpReplyInboxContractTest.releaseCancelsConsumerAndRequeuesHeldDeliveryForRecovery}),
+     * which races this method's own {@code held.remove(target)}: the redelivery can land after the
+     * clear and leave a stale entry behind, so {@link #peek} is no longer reliably empty right after
+     * {@link #release}. Cancelling first closes that consumer, so the requeued message goes back to
+     * the queue for whichever consumer picks it up next (a later {@link #own}), not this one.
+     *
+     * <p><strong>Failure of the requeue is best-effort, not fatal.</strong> {@link #release} runs
+     * during teardown ({@code Fleetd} calls it right after {@code MessageService.abandon}), and a
+     * throw here would abort cleanups the caller depends on — the same argument fleetd #293 settled
+     * for {@code HerdrPeerLauncher.stop()}'s tab-close step. So a failed {@code basicNack} is logged
+     * at WARN, naming the target and delivery tag that leaked, and release proceeds; the delivery
+     * stays unacked on the broker rather than being silently dropped, so it is still recoverable by a
+     * later connection drop even though this release did not manage to requeue it immediately. A
+     * failed {@code basicCancel} still throws, unchanged from before this fix — that failure means
+     * the consumer may still be attached, so best-effort requeue is not attempted underneath it.
+     */
     @Override
     public void release(String target) {
         synchronized (channelLock) {
             String tag = consumerTags.remove(target);
-            held.remove(target);                   // stale delivery tags must not survive release
-            if (tag == null) {
-                return;
+            if (tag != null) {
+                try {
+                    channel.basicCancel(tag);
+                } catch (IOException e) {
+                    throw new IllegalStateException("cannot cancel consumer for " + target, e);
+                }
             }
-            try {
-                channel.basicCancel(tag);
-            } catch (IOException e) {
-                throw new IllegalStateException("cannot cancel consumer for " + target, e);
+            var perTarget = held.remove(target);
+            if (perTarget != null) {
+                synchronized (perTarget) {
+                    for (Held h : perTarget.values()) {
+                        try {
+                            channel.basicNack(h.deliveryTag(), false, true); // requeue, don't drop
+                        } catch (IOException | RuntimeException e) {
+                            // Caught broadly (not just IOException) for the same reason #293 catches
+                            // RuntimeException in HerdrPeerLauncher.stop(): best-effort teardown must
+                            // not be guarded only against the expected failure and bare against any
+                            // other. The message stays unacked on the broker either way — not lost,
+                            // just not proactively requeued — until a connection drop frees it.
+                            log.warn("release({}): could not requeue held delivery (msgId={}, tag={})"
+                                            + " back to the broker — it stays unacked until a connection"
+                                            + " drop frees it: {}",
+                                    target, h.message().msgId(), h.deliveryTag(), e.getMessage());
+                        }
+                    }
+                }
             }
         }
     }

fleetd/src/main/java/dev/ltms/fleet/msg/MessageService.java

@@ -408,10 +408,28 @@ public final class MessageService {
      * as the zero-candidate case does, and let {@link #abandon} apply the eventual recovery
      * deterministically instead.
      *
+     * <p><strong>{@code content} is required (fleetd #302).</strong> Both doors that reach this
+     * method must reject a missing/blank reply the same way, so the check lives here rather than in
+     * either caller: {@code FleetMcp.reply} already refuses a {@code null} content before it ever
+     * calls this method (its own required-arg guard), and no test or production call site anywhere
+     * in the codebase relies on replying with empty content — confirmed by searching every call site
+     * of this method before adding the check, not assumed. Without this guard, a REST {@code
+     * POST /sessions/{id}/reply} whose body omits {@code content} (or a client library that maps a
+     * missing field to {@code ""}) used to reach {@link Rendezvous#resolve} with an empty string,
+     * silently completing the lead's blocking wait with nothing — indistinguishable from a worker
+     * that genuinely replied with nothing, which is worse than a loud failure because it destroys the
+     * information that the reply never arrived.
+     *
+     * @throws IllegalArgumentException if {@code content} is {@code null} or blank — the caller must
+     *         report this as a client error (REST: 400 {@code bad_request}) rather than resolve
+     *         anything
      * @return always {@code true} — the reply resolved a live send, completed a parked ticket, or
      *         was queued
      */
     public boolean reply(String session, String content) {
+        if (content == null || content.isBlank()) {
+            throw new IllegalArgumentException("content is required");
+        }
         if (rendezvous.resolve(session, content)) {
             count(FleetMetrics.REPLIES, "path", "rendezvous");
             return true;                       // a live send took it — unchanged fast path

fleetd/src/main/java/dev/ltms/fleet/rest/FleetApp.java

@@ -9,6 +9,7 @@ import dev.ltms.fleet.auth.Principal;
 import dev.ltms.fleet.guard.GuardException;
 import dev.ltms.fleet.metrics.FleetMetrics;
 import dev.ltms.fleet.metrics.Metrics;
+import dev.ltms.fleet.mcp.FleetMcp;
 import dev.ltms.fleet.herdr.Agent;
 import dev.ltms.fleet.herdr.HerdrClient;
 import dev.ltms.fleet.herdr.HerdrException;
@@ -86,6 +87,12 @@ public final class FleetApp {
     // absent() (the honest "no policy configured" view) for every constructor that does not wire
     // a real one, so existing legacy call sites keep building without knowing this field exists.
     private final Supplier<MemberCredentialPolicyView> memberCredentials;
+    // fleetd #297: the SAME shared sources FleetMcp.profiles/fleet_profiles reads (BackendQuarantine
+    // and BackendOutagePolicy are each one instance for the whole daemon — see Fleetd wiring) so
+    // GET /profiles cannot drift from fleet_profiles about which profile is quarantined/cooling off.
+    // .none() (the honest "feature not wired" view) for every constructor that does not pass one.
+    private final FleetMcp.QuarantineSource quarantine;
+    private final FleetMcp.OutageSource outage;
     private final ObjectMapper mapper = new ObjectMapper();
 
     /**
@@ -146,6 +153,23 @@ public final class FleetApp {
                        MessageService messages, MemberPresence presence,
                        HttpServlet mcpServlet, CallerResolver auth, Metrics metrics,
                        Predicate<String> deliverable, Supplier<MemberCredentialPolicyView> memberCredentials) {
+        this(herdr, memberHerdr, workers, sessions, messages, presence, mcpServlet, auth, metrics,
+                deliverable, memberCredentials, FleetMcp.QuarantineSource.none(), FleetMcp.OutageSource.none());
+    }
+
+    /**
+     * @param quarantine the SAME {@link FleetMcp.QuarantineSource} instance passed to {@code
+     *                   FleetMcp} (fleetd #297), so {@code GET /profiles} reports the identical
+     *                   exhaustion-quarantine facts as {@code fleet_profiles} rather than a second,
+     *                   independently-computed copy
+     * @param outage     the SAME {@link FleetMcp.OutageSource} instance passed to {@code FleetMcp} —
+     *                   see {@code quarantine}; a SEPARATE check from it, never merged in
+     */
+    public FleetApp(HerdrClient herdr, HerdrClient memberHerdr, PeerLauncher workers, SessionManager sessions,
+                       MessageService messages, MemberPresence presence,
+                       HttpServlet mcpServlet, CallerResolver auth, Metrics metrics,
+                       Predicate<String> deliverable, Supplier<MemberCredentialPolicyView> memberCredentials,
+                       FleetMcp.QuarantineSource quarantine, FleetMcp.OutageSource outage) {
         this.herdr = herdr;
         this.memberHerdr = memberHerdr != null ? memberHerdr : herdr;
         this.workers = workers;
@@ -156,6 +180,8 @@ public final class FleetApp {
         this.auth = auth;
         this.metrics = metrics;
         this.memberCredentials = memberCredentials != null ? memberCredentials : MemberCredentialPolicyView::absent;
+        this.quarantine = quarantine != null ? quarantine : FleetMcp.QuarantineSource.none();
+        this.outage = outage != null ? outage : FleetMcp.OutageSource.none();
     }
 
     /** Wire routes onto a fresh, unstarted Javalin instance. Caller starts it. */
@@ -338,8 +364,15 @@ public final class FleetApp {
         if (!allow(ctx, routeAction("GET /agents"), null)) {
             return;
         }
-        ctx.status(200).json(Map.of("agents",
-                workers.list().stream().map(Agent.class::cast).map(FleetApp::view).toList()));
+        try {
+            ctx.status(200).json(Map.of("agents",
+                    workers.list().stream().map(Agent.class::cast).map(FleetApp::view).toList()));
+        } catch (HerdrException e) {
+            // fleetd #297: workers.list() reaches herdr — a transport failure must land in the same
+            // {error, detail} envelope every other failure path here uses, not escape as a bare
+            // exception and leave Javalin's default handling to respond outside the JSON contract.
+            herdrError(ctx, e);
+        }
     }
 
     /** CB-304: bridge-owned roster merged with live herdr status by paneId. */
@@ -347,40 +380,57 @@ public final class FleetApp {
         if (!allow(ctx, routeAction("GET /members"), null)) {
             return;
         }
-        // CB-519: the registry key is a host-unique id, not the pane coordinate — join on terminal.
-        Map<String, Agent> live = workers.list().stream()
-                .map(Agent.class::cast)
-                .filter(a -> a.terminalId() != null)
-                .collect(Collectors.toMap(Agent::terminalId, Function.identity(), (_, b) -> b));
-        // fleetd #209: this REST roster reports agentSessionId via SessionManager.rosterView, so it
-        // uses the resolving roster read (caller-driven, not a timer) rather than the plain one.
-        List<Map<String, Object>> out = sessions.rosterResolved().stream()
-                .map(s -> SessionManager.rosterView(s, live.get(s.terminalId())))
-                .toList();
-        Map<String, Object> body = new LinkedHashMap<>();
-        // fleetd #199: the endpoint became /members in the CB-634 rename but the body key stayed
-        // "workers", so a caller that read "members" saw an empty fleet and reported no members at
-        // all. "members" is the canonical key; "workers" stays as a deprecated alias so an existing
-        // REST consumer keeps working — the out-of-band path a lead falls back to when its MCP mount
-        // drops reads this endpoint. Drop the alias once nothing reads it.
-        body.put("members", out);
-        body.put("workers", out);
-        // CB-586: operator visibility for the refs/wip snapshot store without shelling into the
-        // repo — how many snapshot refs exist and roughly what they cost. Present only once a
-        // worktree session has established the repo, so a never-snapshotted fleet reports nothing.
-        sessions.wipRefs().ifPresent(st -> body.put("wipRefs",
-                Map.of("count", st.count(), "costBytes", st.costBytes())));
-        ctx.status(200).json(body);
+        try {
+            // CB-519: the registry key is a host-unique id, not the pane coordinate — join on terminal.
+            Map<String, Agent> live = workers.list().stream()
+                    .map(Agent.class::cast)
+                    .filter(a -> a.terminalId() != null)
+                    .collect(Collectors.toMap(Agent::terminalId, Function.identity(), (_, b) -> b));
+            // fleetd #209: this REST roster reports agentSessionId via SessionManager.rosterView, so it
+            // uses the resolving roster read (caller-driven, not a timer) rather than the plain one.
+            List<Map<String, Object>> out = sessions.rosterResolved().stream()
+                    .map(s -> SessionManager.rosterView(s, live.get(s.terminalId())))
+                    .toList();
+            Map<String, Object> body = new LinkedHashMap<>();
+            // fleetd #199: the endpoint became /members in the CB-634 rename but the body key stayed
+            // "workers", so a caller that read "members" saw an empty fleet and reported no members at
+            // all. "members" is the canonical key; "workers" stays as a deprecated alias so an existing
+            // REST consumer keeps working — the out-of-band path a lead falls back to when its MCP mount
+            // drops reads this endpoint. Drop the alias once nothing reads it.
+            body.put("members", out);
+            body.put("workers", out);
+            // CB-586: operator visibility for the refs/wip snapshot store without shelling into the
+            // repo — how many snapshot refs exist and roughly what they cost. Present only once a
+            // worktree session has established the repo, so a never-snapshotted fleet reports nothing.
+            sessions.wipRefs().ifPresent(st -> body.put("wipRefs",
+                    Map.of("count", st.count(), "costBytes", st.costBytes())));
+            ctx.status(200).json(body);
+        } catch (HerdrException e) {
+            // fleetd #297: same reasoning as agents() above — this is the out-of-band roster a lead
+            // falls back to when its MCP mount drops, so it must stay inside the JSON error contract
+            // exactly when herdr is briefly unreachable, not escape as a bare exception.
+            herdrError(ctx, e);
+        }
     }
 
-    /** The configured worker profiles and which one a no-argument spawn uses. */
+    /**
+     * The configured worker profiles, which one a no-argument spawn uses, and (fleetd #297) the two
+     * outage states {@code fleet_profiles} already reports: {@code quarantined} (CB-578 stage B —
+     * the backend reported it out of capacity) and {@code coolingOff} (fleetd #201 Unit 5 — the
+     * credential threw repeated non-exhaustion backend errors). Both are read from the SAME shared
+     * {@link FleetMcp.QuarantineSource}/{@link FleetMcp.OutageSource} instances {@code FleetMcp}
+     * reads, never recomputed, so the two doors cannot disagree about which profile is down and why.
+     * Independent checks, so a profile can appear in both maps at once; each map is present only
+     * when at least one profile is in that state.
+     */
     private void profiles(Context ctx) {
         if (!allow(ctx, routeAction("GET /profiles"), null)) {
             return;
         }
-        ctx.status(200).json(Map.of(
-                "profiles", workers.profiles(),
-                "default", workers.defaultProfile() == null ? "" : workers.defaultProfile()));
+        // fleetd #297: ONE body builder, shared with fleet_profiles. Handing both doors the same
+        // QuarantineSource/OutageSource instances stops them reading different facts; rendering
+        // through the same method stops them reporting those facts differently. Both are needed.
+        ctx.status(200).json(FleetMcp.profilesView(workers, quarantine, outage));
     }
 
     /**
@@ -634,7 +684,19 @@ public final class FleetApp {
             ctx.status(400).json(Map.of("error", "bad_request", "detail", "body must be JSON"));
             return;
         }
-        messages.reply(id, content);
+        // fleetd #302: content is required. `.path("content").asText("")` above turns a missing key
+        // into "" rather than throwing, so without this check an empty/blank reply used to reach
+        // messages.reply(...) and silently resolve the lead's waiter — the same class of bug as the
+        // sibling "content is required" guards on sendMessage/askMessage below, except this one wrote
+        // a WRONG value instead of failing loudly. The check lives in MessageService.reply so both
+        // this door and FleetMcp.reply inherit the same rule; this catch only translates it into the
+        // {error, detail} envelope this file uses everywhere else.
+        try {
+            messages.reply(id, content);
+        } catch (IllegalArgumentException e) {
+            ctx.status(400).json(Map.of("error", "bad_request", "detail", e.getMessage()));
+            return;
+        }
         ctx.status(200).json(Map.of("sessionId", id, "delivered", true));
     }
 

fleetd/src/main/java/dev/ltms/fleet/session/SessionManager.java

@@ -864,10 +864,20 @@ public final class SessionManager implements TurnListener {
     }
 
     /**
-     * Gracefully drain all registered sessions on daemon shutdown. For each session that is
-     * {@code BUSY}, poll up to {@code timeoutNanos} for it to leave {@code BUSY}, then release it
-     * regardless. Non-busy sessions are released immediately. A failure releasing one session is
-     * logged and does not abort the rest.
+     * Gracefully drain all registered sessions on daemon shutdown. Non-busy sessions are released
+     * immediately; a {@code BUSY} one is polled until it leaves {@code BUSY}, then released
+     * regardless. A failure releasing one session is logged and does not abort the rest.
+     *
+     * <p>{@code timeoutNanos} is a budget for the WHOLE drain, not a grace period per session: the
+     * deadline is taken once, before the loop. So the first BUSY session can spend all of it, and a
+     * later BUSY one is then released with no wait at all. That is deliberate. This drain is only
+     * one phase of shutdown — {@code Fleetd} closes the message service, the push loop, the
+     * heartbeat, MCP and the router after it — and the whole sequence has to finish inside
+     * launchd's exit window. A per-session grace would let N busy members drain for N * the
+     * timeout, overrun that window, and get the daemon SIGKILLed part-way through; the members not
+     * yet reached would then get no clean release, no preserved-worktree log, and no snapshot.
+     * Cutting one turn short is the cheaper failure, and it is not silent: an abandoned BUSY
+     * session is preserved, snapshotted, and logged at WARN by {@code logPreservedForShutdown}.
      *
      * <p>CB-544: this is a {@link ReleaseCause#SHUTDOWN} release — the worker's pane is stopped
      * (the process must end) but its worktree is preserved and its path logged. Shutdown is never

fleetd/src/test/java/dev/ltms/fleet/herdr/FakeHerdr.java

@@ -7,6 +7,7 @@ import java.util.ArrayList;
 import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.CopyOnWriteArrayList;
 
 /**
@@ -39,8 +40,12 @@ public final class FakeHerdr implements HerdrClient {
     private final Map<String, List<String>> extraTabs = new LinkedHashMap<>();
     private int agentNameTakenFor = 0;
     private int agentPaneBusyFor = 0;
+    private String agentStartErrorCode = null;
     private int workerTabPaneCount = 1;
     private String paneCloseErrorCode = null;
+    private final Map<String, String> paneCloseErrorCodeFor = new ConcurrentHashMap<>();
+    private String tabCloseErrorCode = null;
+    private final Map<String, String> tabCloseErrorCodeFor = new ConcurrentHashMap<>();
     private String agentSendErrorCode = null;
     private boolean noPanes = false;
     private volatile String agentStatus = "idle"; // steady-state agent.get status
@@ -80,18 +85,56 @@ public final class FakeHerdr implements HerdrClient {
         return this;
     }
 
+    /** Make every {@code agent.start} call fail with this herdr error code. */
+    public FakeHerdr agentStartFailsWith(String code) {
+        this.agentStartErrorCode = code;
+        return this;
+    }
+
     /** Make the worker tab (w9:t2) report this many panes in {@code tab.list} (default 1). */
     public FakeHerdr withWorkerTabPaneCount(int n) {
         this.workerTabPaneCount = n;
         return this;
     }
 
-    /** Make {@code pane.close} fail with this herdr error code. */
+    /** Make {@code pane.close} fail with this herdr error code, for every pane. */
     public FakeHerdr paneCloseFailsWith(String code) {
         this.paneCloseErrorCode = code;
         return this;
     }
 
+    /**
+     * Make {@code pane.close} fail with this herdr error code, but only for the given {@code
+     * pane_id} — every other pane's {@code pane.close} still succeeds. Unlike {@link
+     * #paneCloseFailsWith}, which fails every call regardless of which pane it targets, this lets a
+     * test reap/release several sessions at once and make exactly one of them fail to stop, so the
+     * others' teardown can be asserted to proceed normally (fleetd #290).
+     */
+    public FakeHerdr paneCloseFailsForPane(String paneId, String code) {
+        this.paneCloseErrorCodeFor.put(paneId, code);
+        return this;
+    }
+
+    /** Make {@code tab.close} fail with this herdr error code, for every tab. */
+    public FakeHerdr tabCloseFailsWith(String code) {
+        this.tabCloseErrorCode = code;
+        return this;
+    }
+
+    /**
+     * Make {@code tab.close} fail with this herdr error code, but only for the given {@code
+     * tab_id} — every other tab's {@code tab.close} still succeeds. The {@code tab.close}
+     * counterpart to {@link #paneCloseFailsForPane} (fleetd #290): lets a test make exactly one
+     * session's tab teardown fail while proving the rest of {@code stop()} — {@code
+     * releaseZdotdir}, and the caller's worktree removal — still runs (fleetd #293). Named "ForTab"
+     * rather than "ForPane" (unlike its sibling) because {@code tab.close} keys on {@code tab_id},
+     * not a pane id.
+     */
+    public FakeHerdr tabCloseFailsForTab(String tabId, String code) {
+        this.tabCloseErrorCodeFor.put(tabId, code);
+        return this;
+    }
+
     /**
      * Make {@code pane.list} report no panes at all — models a second herdr daemon (CB-185) that
      * simply does not host the pane a {@link PaneLocator} is searching for.
@@ -275,6 +318,10 @@ public final class FakeHerdr implements HerdrClient {
                                             + required + "`", "invalid_request", null);
                         }
                     }
+                    if (agentStartErrorCode != null) {
+                        throw new HerdrException("herdr error [" + agentStartErrorCode + "]: agent.start failed",
+                                agentStartErrorCode, null);
+                    }
                     long starts = calls.stream().filter(c -> c.method().equals("agent.start")).count();
                     if (starts <= agentPaneBusyFor) {
                         throw new HerdrException(
@@ -338,7 +385,17 @@ public final class FakeHerdr implements HerdrClient {
                             .formatted(workerTabPaneCount,
                                     seeded.isEmpty() ? "" : "," + String.join(",", seeded)));
                 }
-                case "tab.close" -> mapper.readTree("{\"type\":\"ok\"}");
+                case "tab.close" -> {
+                    Object tabIdParam = params instanceof Map<?, ?> m ? m.get("tab_id") : null;
+                    String perTabCode = tabIdParam == null ? null
+                            : tabCloseErrorCodeFor.get(String.valueOf(tabIdParam));
+                    String code = perTabCode != null ? perTabCode : tabCloseErrorCode;
+                    if (code != null) {
+                        throw new HerdrException("herdr error [" + code + "]: tab.close failed",
+                                code, null);
+                    }
+                    yield mapper.readTree("{\"type\":\"ok\"}");
+                }
                 case "pane.get" -> mapper.readTree("""
                         {"type":"pane_info","pane":{"pane_id":"w9:pW","workspace_id":"w9",
                           "tab_id":"w9:t2","agent_status":"idle"}}""");
@@ -360,9 +417,13 @@ public final class FakeHerdr implements HerdrClient {
                                   "foreground_processes":[]}}""");
                 }
                 case "pane.close" -> {
-                    if (paneCloseErrorCode != null) {
-                        throw new HerdrException("herdr error [" + paneCloseErrorCode + "]: pane.close failed",
-                                paneCloseErrorCode, null);
+                    Object paneIdParam = params instanceof Map<?, ?> m ? m.get("pane_id") : null;
+                    String perPaneCode = paneIdParam == null ? null
+                            : paneCloseErrorCodeFor.get(String.valueOf(paneIdParam));
+                    String code = perPaneCode != null ? perPaneCode : paneCloseErrorCode;
+                    if (code != null) {
+                        throw new HerdrException("herdr error [" + code + "]: pane.close failed",
+                                code, null);
                     }
                     yield mapper.readTree("{\"type\":\"ok\"}");
                 }

fleetd/src/test/java/dev/ltms/fleet/member/ClaudeCodeLauncherTest.java

@@ -925,6 +925,71 @@ class ClaudeCodeLauncherTest {
         assertTrue(herdr.called("pane.close"), "stop via handle.id() must close the pane");
     }
 
+    /** A tab-placement launcher with {@code memberCredentials policy=allow-list} under a zsh shell — the
+     * combination that makes {@link HerdrPeerLauncher#spawn} generate a real ZDOTDIR, so {@code
+     * releaseZdotdir}'s effect (the directory's deletion) is observable from a test. */
+    private ClaudeCodeLauncher serviceWithAllowList(FakeHerdr herdr) {
+        FleetConfig.Profile cfg = new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
+                List.of("ccs", "ltms-local"), "tab", "fleetd-workers",
+                "worker: {profile} #{n}", null, null, null);
+        Supplier<FleetConfig.MemberCredentials> creds = () -> new FleetConfig.MemberCredentials(
+                FleetConfig.MemberCredentials.POLICY_ALLOW_LIST, List.of(), List.of(), null);
+        Function<String, String> env = name -> "SHELL".equals(name) ? "/bin/zsh" : null;
+        return new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(),
+                env, 0, System::currentTimeMillis, () -> { }, null, creds);
+    }
+
+    /**
+     * fleetd #293: {@code stop()} used to run {@code spaces.closeTab} bare — any non-{@code
+     * *_not_found} herdr error propagated straight out of {@code stop()}, skipping {@code
+     * releaseZdotdir} entirely (the pane was already closed by that point, so the tab-close failure
+     * is cosmetic, not a real teardown failure). Proves both halves of the fix: {@code stop()} no
+     * longer throws for this failure, and {@code releaseZdotdir} still runs — observed here by the
+     * generated ZDOTDIR actually being deleted, since {@code releaseZdotdir}'s last line is {@code
+     * EnvAllowListScrub.deleteRecursively(dir)}.
+     */
+    @Test
+    @SuppressWarnings("unchecked")
+    void stopStillReleasesZdotdirWhenCloseTabFailsWithANonNotFoundCode() {
+        FakeHerdr herdr = new FakeHerdr();
+        ClaudeCodeLauncher svc = serviceWithAllowList(herdr);
+        PeerHandle handle = svc.spawn(new SpawnRequest(null, null, null));
+        Map<String, Object> tabCreateParams = (Map<String, Object>) herdr.lastCall("tab.create").params();
+        Map<String, String> tabEnv = (Map<String, String>) tabCreateParams.get("env");
+        String zdotdir = tabEnv.get("ZDOTDIR");
+        assertNotNull(zdotdir, "policy=allow-list under a zsh shell must have generated a ZDOTDIR: " + tabEnv);
+        Path dir = Path.of(zdotdir);
+        assertTrue(Files.isDirectory(dir), "the generated ZDOTDIR must exist before stop(): " + dir);
+        herdr.tabCloseFailsForTab("w9:t2", "internal_error");
+
+        Logger logger = (Logger) LoggerFactory.getLogger(HerdrPeerLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            assertDoesNotThrow(() -> svc.stop(handle.id()),
+                    "fleetd #293: a failing tab.close is cosmetic — it must not propagate out of stop()");
+        } finally {
+            logger.detachAppender(appender);
+        }
+
+        assertTrue(herdr.called("tab.close"), "tab.close was still attempted");
+        assertFalse(Files.exists(dir),
+                "releaseZdotdir must still run and delete the generated ZDOTDIR despite the tab.close "
+                        + "failure: " + dir);
+        String warn = appender.list.stream()
+                .filter(e -> e.getLevel().equals(Level.WARN))
+                .map(ILoggingEvent::getFormattedMessage)
+                .filter(m -> m.contains("tab.close") && m.contains("w9:t2"))
+                .findFirst()
+                .orElse(null);
+        assertNotNull(warn, "the failing tab.close must be logged at WARN naming the tab id — a "
+                + "silently swallowed failure with no message is not an improvement. Log lines: "
+                + appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList());
+    }
+
     // --- CB-519: host-unique id, decoupled from the pane coordinate ------------------------------
 
     @Test
@@ -1100,7 +1165,8 @@ class ClaudeCodeLauncherTest {
     @Test
     void spawnLetsAnUnrelatedHerdrErrorPropagateUnchanged() {
         // Fix 1 must only special-case a "*_not_found" answer. Any other herdr failure keeps
-        // propagating as-is — this gate does not know how to recover from it.
+        // propagating as-is — this gate does not know how to recover from it. The pane still needs
+        // closing because spawn throws before it can return the pane id to a caller that could stop it.
         FakeHerdr herdr = new FakeHerdr();
         herdr.agentStatus("unknown");
         herdr.agentGetFailsWithAfter(0, "internal_error");
@@ -1117,8 +1183,27 @@ class ClaudeCodeLauncherTest {
                 () -> svc.spawn(new SpawnRequest(null, null, null)));
 
         assertEquals("internal_error", ex.code());
-        assertEquals(0, paneCloseCount(herdr, "w9:pRoot_1"),
-                "an error this gate does not recognize is not this gate's teardown to run");
+        assertEquals(1, paneCloseCount(herdr, "w9:pRoot_1"),
+                "the unchanged error leaves spawn without a pane id, so this gate closes its orphaned pane");
+    }
+
+    @Test
+    void panePlacementClosesTheSplitPaneWhenAgentStartFails() {
+        FakeHerdr herdr = new FakeHerdr().agentStartFailsWith("internal_error");
+        FleetConfig.Profile cfg = new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
+                List.of("claude"), "pane", "fleetd-workers", "w #{n}", null, null, null);
+        ClaudeCodeLauncher svc = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null);
+
+        dev.ltms.fleet.herdr.HerdrException ex = assertThrows(
+                dev.ltms.fleet.herdr.HerdrException.class,
+                () -> svc.spawn(new SpawnRequest(null, null, null)));
+
+        assertEquals("internal_error", ex.code(), "agent.start failure propagates unchanged");
+        assertEquals(1, paneCloseCount(herdr, "w1:pSplit"),
+                "the pane split for a peer that never starts is closed instead of left orphaned");
     }
 
     // --- fleetd #176 fix 2: corroborated UNKNOWN refinement --------------------------------------

fleetd/src/test/java/dev/ltms/fleet/msg/AmqpReplyInboxContractTest.java

@@ -154,7 +154,7 @@ class AmqpReplyInboxContractTest {
     }
 
     @Test
-    void releaseCancelsConsumerAndClearsHeld() throws Exception {
+    void releaseCancelsConsumerAndRequeuesHeldDeliveryForRecovery() throws Exception {
         String target = "worker-release-" + System.nanoTime();
         try (AmqpReplyInbox inbox = AmqpReplyInbox.open(uri())) {
             inbox.own(target);
@@ -164,6 +164,22 @@ class AmqpReplyInboxContractTest {
             inbox.release(target);
             assertTrue(inbox.peek(target).isEmpty(),
                     "release clears the local held snapshot");
+
+            // fleetd #298: release() must not just drop the local record — the broker delivery was
+            // never acked, so cancelling the consumer alone leaves it unacked-but-orphaned on the
+            // still-open channel unless release() nacks it back with requeue=true. Prove the message
+            // is genuinely recoverable, not merely absent from peek: re-own the same target and
+            // confirm the broker redelivers it to the fresh consumer.
+            inbox.own(target);
+            List<ReplyInbox.InboxMessage> recovered = awaitPeek(inbox, target);
+            assertEquals(1, recovered.size(),
+                    "a reply held (but undrained) at release() time must still be recoverable — "
+                            + "release() must requeue it, not silently drop it while the broker still "
+                            + "considers it outstanding");
+            assertEquals("m1", recovered.getFirst().msgId());
+            assertEquals("release me", recovered.getFirst().content());
+
+            inbox.ack(target, "m1");
         }
     }
 

fleetd/src/test/java/dev/ltms/fleet/rest/FleetAppTest.java

@@ -19,6 +19,10 @@ import dev.ltms.fleet.session.SessionManager;
 import dev.ltms.fleet.session.Worktrees;
 import dev.ltms.fleet.member.ClaudeCodeLauncher;
 import dev.ltms.fleet.member.CompositePeerLauncher;
+import dev.ltms.fleet.member.MemberCredentialPolicyView;
+import dev.ltms.fleet.mcp.FleetMcp;
+import dev.ltms.fleet.placement.BackendOutagePolicy;
+import dev.ltms.fleet.placement.BackendQuarantine;
 import dev.ltms.fleet.placement.PlacementPolicies;
 import io.javalin.Javalin;
 import org.junit.jupiter.api.AfterEach;
@@ -32,6 +36,7 @@ import java.util.List;
 import java.util.Map;
 import java.util.Set;
 import java.util.UUID;
+import java.util.concurrent.TimeUnit;
 import java.util.function.Predicate;
 
 import static org.junit.jupiter.api.Assertions.*;
@@ -70,6 +75,19 @@ class FleetAppTest {
 
     private int start(FakeHerdr herdr, String workerBaseUrl, Set<String> allow, String placement,
                       Worktrees worktrees, Predicate<String> deliverable) {
+        return start(herdr, workerBaseUrl, allow, placement, worktrees, deliverable,
+                FleetMcp.QuarantineSource.none(), FleetMcp.OutageSource.none());
+    }
+
+    /**
+     * fleetd #297: same wiring as above, plus the two SAME shared sources {@code GET /profiles}
+     * must read — lets a test prove the quarantined/coolingOff facts it reports come from a real
+     * {@link dev.ltms.fleet.placement.BackendQuarantine}/{@link
+     * dev.ltms.fleet.placement.BackendOutagePolicy}, exactly like {@code fleet_profiles}'s own tests.
+     */
+    private int start(FakeHerdr herdr, String workerBaseUrl, Set<String> allow, String placement,
+                      Worktrees worktrees, Predicate<String> deliverable,
+                      FleetMcp.QuarantineSource quarantine, FleetMcp.OutageSource outage) {
         FleetConfig.Profile wcfg = new FleetConfig.Profile(
                 "ltms-local", workerBaseUrl, "coder", null, "FLEETD_WORKER_TOKEN", null,
                 placement, "fleet", "worker: {profile} #{n}", null, null, null);
@@ -90,8 +108,9 @@ class FleetAppTest {
         // it directly so the inbox contract holds for those endpoints.
         inbox.own("term_a");
         MessageService messages = new MessageService(agents, injector, rendezvous, inbox);
-        app = new FleetApp(herdr, workers, sessions, messages, this.presence, null,
-                null, null, id -> this.presence.isPresent(id) || deliverable.test(id))
+        app = new FleetApp(herdr, herdr, workers, sessions, messages, this.presence, null,
+                null, null, id -> this.presence.isPresent(id) || deliverable.test(id),
+                MemberCredentialPolicyView::absent, quarantine, outage)
                 .build().start("127.0.0.1", 0);
         return app.port();
     }
@@ -164,6 +183,22 @@ class FleetAppTest {
         assertEquals("idle", agents.get(0).get("status").asText());
     }
 
+    /**
+     * fleetd #297 gap 1: {@code workers.list()} reaches herdr, and a transport failure there must
+     * land in the same {@code {error, detail}} envelope every other failure path in this file uses
+     * (see {@code herdrError}), not escape as a bare exception outside the JSON contract.
+     */
+    @Test
+    void agentsMapsAHerdrFailureToTheJsonErrorEnvelope() throws Exception {
+        FakeHerdr down = new FakeHerdr().healthy(false);
+        int port = start(down, "http://gx00.gw:8000", Set.of("gx00.gw"));
+        HttpResponse<String> res = req(port, "GET", "/agents");
+        assertEquals(502, res.statusCode(), res.body());
+        JsonNode body = mapper.readTree(res.body());
+        assertEquals("herdr_error", body.get("error").asText());
+        assertTrue(body.has("detail"), res.body());
+    }
+
     @Test
     void spawnWorkerLandsInOwnTabInWorkerSpaceAndInjectsBaseUrl() throws Exception {
         FakeHerdr herdr = new FakeHerdr();
@@ -203,6 +238,42 @@ class FleetAppTest {
         JsonNode body = mapper.readTree(req(port, "GET", "/profiles").body());
         assertEquals("ltms-local", body.get("default").asText());
         assertEquals("ltms-local", body.get("profiles").get(0).asText());
+        assertFalse(body.has("quarantined"), "nothing is quarantined, so the key is omitted: " + body);
+        assertFalse(body.has("coolingOff"), "nothing is cooling off, so the key is omitted: " + body);
+    }
+
+    /**
+     * fleetd #297 gap 2: {@code GET /profiles} must report the same two outage states {@code
+     * fleet_profiles} does — CB-578 stage B exhaustion quarantine and fleetd #201 Unit 5 cool-off —
+     * reading the SAME shared {@link BackendQuarantine}/{@link BackendOutagePolicy} instances rather
+     * than recomputing them. The two checks are independent, and this profile is deliberately put in
+     * both states at once, matching {@code FleetMcpTest}'s own coverage of that overlap.
+     */
+    @Test
+    void profilesReportsQuarantineAndCoolingOffFromTheSameSharedSources() throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        BackendQuarantine quarantine = new BackendQuarantine(() -> 0L, TimeUnit.MINUTES.toNanos(30));
+        quarantine.quarantine("shared-openai");
+        FleetMcp.QuarantineSource quarantineSource = new FleetMcp.QuarantineSource(
+                profile -> "ltms-local".equals(profile) ? "shared-openai" : null, quarantine);
+        BackendOutagePolicy outagePolicy = new BackendOutagePolicy(() -> 0L);
+        outagePolicy.record("shared-openai", "t1", "API Error: rate limited");
+        outagePolicy.record("shared-openai", "t2", "API Error: rate limited"); // 2nd distinct target starts the incident
+        FleetMcp.OutageSource outageSource = new FleetMcp.OutageSource(
+                profile -> "ltms-local".equals(profile) ? "shared-openai" : null, outagePolicy);
+        int port = start(herdr, "http://gx00.gw:8000", Set.of("gx00.gw"), "tab", new GitWorktrees(),
+                ignored -> false, quarantineSource, outageSource);
+
+        JsonNode body = mapper.readTree(req(port, "GET", "/profiles").body());
+        assertTrue(body.has("quarantined"), body.toString());
+        assertEquals("shared-openai",
+                body.get("quarantined").get("ltms-local").get("credentialId").asText());
+        assertEquals(1800,
+                body.get("quarantined").get("ltms-local").get("quarantinedForSeconds").asLong());
+        assertTrue(body.has("coolingOff"), body.toString());
+        assertEquals("shared-openai",
+                body.get("coolingOff").get("ltms-local").get("credentialId").asText());
+        assertEquals(60, body.get("coolingOff").get("ltms-local").get("coolingOffForSeconds").asLong());
     }
 
     @Test
@@ -239,6 +310,23 @@ class FleetAppTest {
                 "liveStatus is unknown when herdr has no matching pane");
     }
 
+    /**
+     * fleetd #297 gap 1: same reasoning as {@code agentsMapsAHerdrFailureToTheJsonErrorEnvelope} —
+     * {@code GET /members} is the endpoint's own comment names as "the out-of-band path a lead falls
+     * back to when its MCP mount drops", so it must stay inside the {@code {error, detail}} envelope
+     * exactly when herdr is briefly unreachable.
+     */
+    @Test
+    void membersMapsAHerdrFailureToTheJsonErrorEnvelope() throws Exception {
+        FakeHerdr down = new FakeHerdr().healthy(false);
+        int port = start(down, "http://gx00.gw:8000", Set.of("gx00.gw"));
+        HttpResponse<String> res = req(port, "GET", "/members");
+        assertEquals(502, res.statusCode(), res.body());
+        JsonNode body = mapper.readTree(res.body());
+        assertEquals("herdr_error", body.get("error").asText());
+        assertTrue(body.has("detail"), res.body());
+    }
+
     @Test
     void spawnWithACwdParamRootsTheWorkerThere() throws Exception {
         FakeHerdr herdr = new FakeHerdr();
@@ -444,6 +532,53 @@ class FleetAppTest {
         assertEquals("orphan", body.get("replies").get(0).get("content").asText());
     }
 
+    @Test
+    void replyWithMissingContentIsRejectedAndDoesNotResolveTheWaiter() throws Exception {
+        // fleetd #302: `.path("content").asText("")` used to turn a missing "content" key into an
+        // empty string that reached rendezvous.resolve, silently completing the lead's blocking wait
+        // with nothing. Prove the fix two ways: the bad call is rejected with 400, AND the real send
+        // it would have wrongly resolved is still open afterwards — a real reply completes it.
+        FakeHerdr herdr = new FakeHerdr().agentStatus("idle");
+        int port = start(herdr, "http://gx00.gw:8000", Set.of("gx00.gw"));
+
+        var send = java.util.concurrent.CompletableFuture.supplyAsync(() -> {
+            try { return postMessage(port, "{\"content\":\"review this\",\"timeoutMs\":4000}"); }
+            catch (Exception e) { throw new RuntimeException(e); }
+        });
+
+        Thread.sleep(200); // let the background send open its rendezvous waiter
+
+        HttpResponse<String> badReply = postJson(port, "/sessions/term_a/reply", "{}");
+        assertEquals(400, badReply.statusCode());
+        JsonNode err = mapper.readTree(badReply.body());
+        assertEquals("bad_request", err.get("error").asText());
+        assertTrue(err.has("detail"));
+
+        // The waiter must still be open — a real reply now completes the ORIGINAL send.
+        HttpResponse<String> goodReply = postJson(port, "/sessions/term_a/reply", "{\"content\":\"LGTM ship it\"}");
+        assertEquals(200, goodReply.statusCode());
+
+        HttpResponse<String> res = send.get(6, java.util.concurrent.TimeUnit.SECONDS);
+        assertEquals(200, res.statusCode());
+        assertEquals("LGTM ship it", mapper.readTree(res.body()).get("reply").asText());
+    }
+
+    @Test
+    void replyWithEmptyOrWhitespaceContentIsRejectedSameAsMissing() throws Exception {
+        // fleetd #302 sibling case: present-but-blank content is treated the same as a missing key —
+        // FleetMcp's own required-content guard (fleet_reply's "content is required") makes no
+        // distinction between the two either, so diverging here would be a new asymmetry.
+        int port = startHealthy();
+
+        HttpResponse<String> empty = postJson(port, "/sessions/term_a/reply", "{\"content\":\"\"}");
+        assertEquals(400, empty.statusCode());
+        assertEquals("bad_request", mapper.readTree(empty.body()).get("error").asText());
+
+        HttpResponse<String> whitespace = postJson(port, "/sessions/term_a/reply", "{\"content\":\"   \"}");
+        assertEquals(400, whitespace.statusCode());
+        assertEquals("bad_request", mapper.readTree(whitespace.body()).get("error").asText());
+    }
+
     @Test
     void drainRepliesReturnsEmptyForNoReplies() throws Exception {
         int port = startHealthy();

fleetd/src/test/java/dev/ltms/fleet/session/SessionManagerTest.java

@@ -1069,6 +1069,79 @@ class SessionManagerTest {
         assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_3"), "the third pane is stopped");
     }
 
+    /**
+     * fleetd #290: the #283 fix above closed the one trigger this suite used for {@code
+     * reapIdle}'s own per-session try/catch (CB-581) — a worktree-removal failure is now caught
+     * and logged inside {@code release()} itself, so it never reaches {@code reapIdle}'s guard at
+     * all. This test restores coverage of that guard using the trigger the ticket names: {@code
+     * release()} calls {@code launcher.stop(paneId)} with no try/catch around it, so a failing
+     * {@code pane.close} propagates straight out of {@code release()} uncaught. {@link
+     * FakeHerdr#paneCloseFailsForPane} (added for this ticket) makes exactly the middle session's
+     * stop fail, while the other two still succeed, so this proves {@code reapIdle} keeps reaping
+     * the rest of the roster rather than aborting the whole pass.
+     */
+    @Test
+    void reapIdleSurvivesOneSessionWhoseLauncherStopFails() {
+        long[] clock = {0};
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees();
+        SessionManager sessions = sessionManager(herdr, worktrees, () -> clock[0]);
+        MemberSession a = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-290a", null));
+        MemberSession b = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-290b", null));
+        MemberSession c = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-290c", null));
+        sessions.asPresence().markPresent(a.terminalId());
+        sessions.asPresence().markPresent(b.terminalId());
+        sessions.asPresence().markPresent(c.terminalId());
+        // Only the middle session's herdr pane fails to close — a and c stop normally. This is the
+        // trigger reapIdle's own guard is for, now that #283 closed the worktree-removal trigger.
+        herdr.paneCloseFailsForPane("w9:pRoot_2", "internal_error");
+
+        LoggerContext ctx = (LoggerContext) LoggerFactory.getILoggerFactory();
+        ch.qos.logback.classic.Logger sessionLog =
+                (ch.qos.logback.classic.Logger) LoggerFactory.getLogger(SessionManager.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.setContext(ctx);
+        appender.start();
+        sessionLog.addAppender(appender);
+        sessionLog.setLevel(Level.WARN);
+        int reaped;
+        try {
+            clock[0] = 100;
+            reaped = sessions.reapIdle(10);
+
+            String warn = appender.list.stream()
+                    .filter(e -> e.getLevel().equals(Level.WARN))
+                    .map(ILoggingEvent::getFormattedMessage)
+                    .filter(m -> m.contains("reap failed") && m.contains(b.paneId()))
+                    .findFirst()
+                    .orElse("no reap-failed WARN logged for the failing session");
+            assertTrue(warn.contains(b.terminalId()), "the WARN names the failed session's terminal: " + warn);
+        } finally {
+            sessionLog.detachAppender(appender);
+        }
+
+        assertEquals(2, reaped,
+                "the middle session's launcher.stop failure is not counted as reaped, but must not "
+                        + "abort reaping the other two");
+        assertTrue(sessions.get(a.paneId()).isEmpty(), "the first session is still released");
+        assertTrue(sessions.get(c.paneId()).isEmpty(),
+                "the third session is still reached and released — proves the pass did not abort "
+                        + "when the middle session's release() threw");
+        assertTrue(sessions.get(b.paneId()).isEmpty(),
+                "the middle session is still deregistered — release() removes it from the registry "
+                        + "before launcher.stop() runs, regardless of whether stop() then throws");
+        assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_1"), "the first pane is stopped");
+        assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_2"),
+                "the middle pane's stop was attempted, even though it failed");
+        assertEquals(1, paneCloseCallsFor(herdr, "w9:pRoot_3"), "the third pane is stopped");
+        assertEquals(List.of(a.worktree(), c.worktree()), worktrees.removeCalls().stream().sorted().toList(),
+                "the middle session's worktree removal never runs — release() throws before reaching "
+                        + "it — while the other two, unaffected, still have theirs removed");
+    }
+
     @Test
     void unchangedRegressionCleanCompletedReleaseStillRemovesTheWorktree() {
         FakeHerdr herdr = new FakeHerdr();
@@ -1083,6 +1156,41 @@ class SessionManagerTest {
                 "COMPLETED release of a clean worktree still removes it");
     }
 
+    /**
+     * fleetd #293: {@code HerdrPeerLauncher.stop()} used to run {@code spaces.closeTab} bare — a
+     * failing {@code tab.close} (any code other than {@code *_not_found}) propagated straight out
+     * of {@code stop()}. {@code SessionManager.release} calls {@code launcher.stop(paneId)} with
+     * no try/catch (fleetd #283 wrapped the WORKTREE-removal step further down, not this one), so
+     * the throw happened <em>before</em> that worktree-removal step ever ran — and by then {@code
+     * registry.remove(paneId)} had already run, so a second {@code stop} is a no-op: the worktree
+     * leaked with no retry path. The pane itself is already closed by the time {@code tab.close}
+     * runs, so its failure is cosmetic workspace tidying, not a real teardown failure. The fix
+     * wraps {@code closeTab} inside {@code stop()} so it no longer throws for this reason; this
+     * test proves both halves at once: {@code release()} does not throw, and it still removes the
+     * worktree.
+     */
+    @Test
+    void releaseStillRemovesTheWorktreeWhenCloseTabFails() {
+        FakeHerdr herdr = new FakeHerdr();
+        RecordingWorktrees worktrees = new RecordingWorktrees();
+        SessionManager sessions = sessionManager(herdr, worktrees);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-293a", null));
+        // FakeHerdr's pane.get always answers with tab_id "w9:t2" for a tab-placement spawn.
+        herdr.tabCloseFailsForTab("w9:t2", "internal_error");
+
+        assertDoesNotThrow(() -> sessions.release(s.paneId()),
+                "a failing tab.close is cosmetic (the pane is already closed by then) — it must not "
+                        + "propagate out of release()");
+
+        assertTrue(herdr.called("tab.close"), "tab.close was still attempted");
+        assertEquals(List.of(s.worktree()), worktrees.removeCalls(),
+                "release() must still remove the worktree even though tab.close failed — this is "
+                        + "the leak fleetd #293 reports: before the fix, release() never reached this "
+                        + "step at all");
+        assertTrue(sessions.get(s.paneId()).isEmpty(), "the session is still deregistered");
+    }
+
     @Test
     void unchangedRegressionDirtyCompletedReleaseStillPreservesTheWorktree() {
         FakeHerdr herdr = new FakeHerdr();