CB-637: relabel lead-comms (was colliding CB-635) + document coordId route in the CLAUDE.md tool table

The lead-to-lead wiring shipped with CB-635 in its comments, but CB-635 is
already the broker.uriEnv / unreachable-broker work. Relabel the mailbox +
fleet_send{coordId} + receive loop to CB-637 so a ticket number names one
feature. Add the cross-host peer-lead row to the primary intent->tool table
(kept byte-identical with the wiki template).

CLAUDE.md

@@ -112,7 +112,8 @@ the merge — and merging on a reviewer's word is delegating it by proxy.
 | Delegate (blocking) | `fleet_send{sessionId, content}` |
 | Delegate (long task) | `fleet_send{sessionId, content, wait:false}` → ticket → `fleet_poll{ticket}` |
 | Answer a member's `fleet_ask` | `fleet_send{turnId, content}` — **not** `sessionId` |
-| Message a **peer lead** | `fleet_send{sessionId: <their terminal>, content}` — `fleet_list` → `leads` reports it. Coordination only, **never** a task |
+| Message a **peer lead** on this host | `fleet_send{sessionId: <their terminal>, content}` — `fleet_list` → `leads` reports it. Coordination only, **never** a task |
+| Message a **peer lead** on another daemon or host | `fleet_send{coordId: <their coord-id>, content}` — needs a `coordinator:` block; your own coord-id is in `fleet_list`. Coordination only, **never** a task |
 | Answer a peer lead that messaged you | `fleet_reply{content}` — the one case a lead replies |
 | Collect a held reply | `fleet_poll{target}` · then `fleet_ack{target, msgId}` |
 | Tear down a member | `fleet_stop{paneId}` |

bridged/fleetd.example.yaml

@@ -140,6 +140,18 @@ herdrSocket: ~/.config/herdr/herdr.sock
 #                      e.g. `env DISPLAY=:10.0 idea {dir}`. Best-effort: a failure is logged, never
 #                      fails the spawn. Omit to open the member's module by hand. There is no close
 #                      half yet — an opened module stays open until the operator closes it.
+#   autoCompactWindow → opt-in, default off. A bounded token window that forces a spawned member to
+#                      compact its context instead of running on the backend's own default and dying
+#                      mid-turn (losing its fleet_reply — the whole point of the turn — with it).
+#                      Validated at config load to [100000, 1000000] — the band Claude Code's own
+#                      --autocompact flag accepts.
+#                      CROSS-BACKEND SEMANTICS DIFFER: on claude-code this is a launch-time
+#                      `--autocompact <tokens>` flag — the member compacts AT this window. opencode
+#                      has no equivalent flag (it only forces `compaction.auto: true`, unconditionally,
+#                      already), so this is instead applied as the model's `limit.context` in the
+#                      generated opencode.json — the member compacts WITHIN this window, not exactly
+#                      at it — and only when this profile's `model:` is in `provider/model` form; if it
+#                      isn't, bridged logs a WARN naming the profile rather than silently doing nothing.
 #   tokenEnv         → host env var holding the worker's auth token (value never stored in config);
 #                      omit for a backend that needs no token (e.g. a local ollama).
 #   cwd              → pin this profile's working directory (CB-112). Omit to inherit the primary's
@@ -256,6 +268,7 @@ profiles:
     # ideMcpUrl: http://127.0.0.1:29170/index-mcp/streamable-http   # opt-in (CB-634): IDE code intelligence, pinned to the worktree
     # ideProjectDir: bridged          # CB-634: module dir the IDE opens + the overlay pins (this repo's pom is in bridged/)
     # ideOpenCommand: env DISPLAY=:10.0 idea {dir}   # CB-634 auto-open: opens {dir} in the IDE at spawn; omit to open by hand
+    # autoCompactWindow: 250000   # opt-in: bound member context; claude-code compacts AT this, opencode within it (model limit.context)
   gx11:                              # a second backend, so `placement: weighted` has a choice
     baseUrl: http://gx01.gw:8000     # self-hosted; ccs handles the model + token
     placement: tab

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

@@ -31,6 +31,9 @@ import dev.ltms.fleet.mcp.LsofPeerPidLookup;
 import dev.ltms.fleet.mcp.LsofProcessCwdLookup;
 import dev.ltms.fleet.msg.AmqpReplyInbox;
 import dev.ltms.fleet.msg.InMemoryReplyInbox;
+import dev.ltms.fleet.msg.LeadChannel;
+import dev.ltms.fleet.msg.LeadCoordLoop;
+import dev.ltms.fleet.msg.LeadMailbox;
 import dev.ltms.fleet.msg.MessageService;
 import dev.ltms.fleet.msg.Rendezvous;
 import dev.ltms.fleet.msg.ReplyInbox;
@@ -60,6 +63,7 @@ import java.util.Map;
 import java.util.Objects;
 import java.util.Set;
 import java.util.concurrent.Executors;
+import java.util.concurrent.ScheduledExecutorService;
 import java.util.concurrent.TimeUnit;
 import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.Function;
@@ -79,6 +83,14 @@ public final class Fleetd {
 
     /** CB-504: how long to wait at startup for herdr's socket before serving degraded. */
     private static final long HERDR_WAIT_SECONDS = 30;
+    /**
+     * CB-637: how often the lead coordination loop looks for peer messages. A few seconds — slow
+     * enough that an idle fleet is not polling a broker in a tight loop, fast enough that a peer
+     * lead's message is not left sitting once the local lead reaches a turn boundary. The mailbox
+     * pushes into the loop's held set on its own consumer thread, so this interval bounds only the
+     * pane delivery, never the receive.
+     */
+    private static final long LEAD_COORD_INTERVAL_MS = 3_000L;
     private static final long HERDR_WAIT_POLL_MILLIS = 500;
 
     /**
@@ -375,6 +387,12 @@ public final class Fleetd {
         // unusable), bridged stays soft-state on the in-memory inbox. The AMQP inbox owns a broker
         // connection, so keep the reference to close it in the ordered shutdown hook.
         final ReplyInbox replyInbox = selectReplyInbox(cfg.broker(), System.getenv(), AmqpReplyInbox::open);
+        // CB-637: this daemon's lead-to-lead mailbox on the SHARED coordination vhost — a separate
+        // broker from the reply inbox by design (see FleetConfig.Coordinator). Absent a coordinator:
+        // block this is null and every lead path below is simply not wired, which is exactly the
+        // behaviour before this ticket. It owns a broker connection, so keep the reference for the
+        // ordered shutdown hook.
+        final LeadMailbox leadMailbox = openLeadMailbox(cfg.coordinator(), System.getenv(), LeadMailbox::open);
         // CB-307: learn the primary's terminal from orchestration tool calls (or pin from config).
         // The pin also feeds CallerResolver below: a primary running inside a herdr pane would
         // otherwise resolve as a worker and be refused every orchestration tool.
@@ -503,7 +521,27 @@ public final class Fleetd {
                 new FleetMcp.QuarantineSource(profile -> {
                     var configured = config.get().profiles().get(profile);
                     return configured == null ? null : configured.effectiveCredentialId();
-                }, quarantine));
+                }, quarantine),
+                leadMailbox);
+
+        // CB-637: the receive half. Only constructed when a lead mailbox actually opened — with no
+        // coordinator (or an unreachable one) there is nothing to deliver, so no scheduler is
+        // created and no thread runs. It reads the SAME live lead supplier the injector's
+        // deliverability gate does, so a lead found by the tab scan after startup is reachable
+        // without a restart.
+        final LeadCoordLoop leadCoordLoop;
+        final ScheduledExecutorService leadCoordSchedulerRef;
+        if (leadMailbox != null) {
+            var leadCoordScheduler = Executors.newSingleThreadScheduledExecutor(r ->
+                    Thread.ofVirtual().name("bridge-leadcoord-").unstarted(r));
+            leadCoordLoop = new LeadCoordLoop(leadMailbox, agents, leads, leadCoordScheduler,
+                    LEAD_COORD_INTERVAL_MS);
+            leadCoordLoop.start();
+            leadCoordSchedulerRef = leadCoordScheduler;
+        } else {
+            leadCoordLoop = null;
+            leadCoordSchedulerRef = null;
+        }
 
         // CB-559: opt-in config reload. With no `configReload:` block nothing is constructed, so an
         // upgraded daemon behaves exactly as before — the file is read once at boot and never again.
@@ -524,6 +562,8 @@ public final class Fleetd {
             messages.close();
             pushLoop.close();
             if (heartbeat != null) heartbeat.close(); // CB-551: stop the idle-lead heartbeat scheduler
+            if (leadCoordLoop != null) leadCoordLoop.close(); // CB-637: stop delivering peer-lead messages
+            if (leadCoordSchedulerRef != null) leadCoordSchedulerRef.shutdownNow();
             if (healthMonitor != null) healthMonitor.stop();
             if (configWatcher != null) configWatcher.stop(); // CB-559: stop polling the config file
             mcp.close();
@@ -536,6 +576,15 @@ public final class Fleetd {
                     log.debug("reply inbox close: {}", e.toString());
                 }
             }
+            // CB-637: the coordination connection goes with it — after the loop that reads it has
+            // stopped, so no tick can be mid-ack against a closed channel.
+            if (leadMailbox != null) {
+                try {
+                    leadMailbox.close();
+                } catch (Exception e) {
+                    log.debug("lead mailbox close: {}", e.toString());
+                }
+            }
             herdr.close();
         }));
 
@@ -575,6 +624,66 @@ public final class Fleetd {
         ReplyInbox open(String uri, int prefetch);
     }
 
+    /** Injection seam for {@link #openLeadMailbox}: production binds {@link LeadMailbox#open}. */
+    @FunctionalInterface
+    interface LeadMailboxOpener {
+        LeadMailbox open(String uri, String selfCoordId, int prefetch);
+    }
+
+    /**
+     * CB-637: open this daemon's lead-to-lead mailbox, or return {@code null} to leave the feature
+     * off. Package-private and env-injected for the same reason as {@link #selectReplyInbox}: the
+     * selection is then testable without a broker or a mutable process environment.
+     *
+     * <p>Every "off" path returns {@code null}, and each says why at the level it deserves:
+     *
+     * <ul>
+     *   <li>no {@code coordinator:} block — silent. Lead coordination is opt-in; an operator who
+     *       never configured it does not need to be told it is off on every boot.</li>
+     *   <li>a block whose {@code uriEnv} does not resolve — INFO, the same "you moved to the secret
+     *       store and the variable is not there" case {@code selectReplyInbox} warns about.</li>
+     *   <li>a configured broker but no {@code selfId} — WARN. This one is a half-finished config: a
+     *       mailbox is named after the coord-id that owns it, so with no id there is no queue to own
+     *       and no {@code from} to send as. Loud, because the operator plainly intended the feature.</li>
+     *   <li>the broker refuses at boot — WARN, and carry on. Mirrors {@code openAmqpOrFallback}: a
+     *       coordination broker that is down must never take a whole fleet's daemon with it, and the
+     *       fleet still works exactly as it did before this feature existed.</li>
+     * </ul>
+     */
+    static LeadMailbox openLeadMailbox(FleetConfig.Coordinator coordinator, Map<String, String> env,
+                                       LeadMailboxOpener opener) {
+        if (coordinator == null) {
+            return null;                       // opt-in: nothing configured, nothing to say
+        }
+        String uri = coordinator.effectiveUri(env);
+        if (uri == null) {
+            log.info("lead coordination: OFF — coordinator{} has no usable broker uri",
+                    coordinator.uriEnv() == null ? "" : ".uriEnv=" + coordinator.uriEnv());
+            return null;
+        }
+        if (coordinator.selfId() == null || coordinator.selfId().isBlank()) {
+            log.warn("coordinator.selfId is unset — lead coordination is OFF. A lead mailbox is the "
+                    + "queue named after the coord-id that owns it, so with no id there is nothing to "
+                    + "own and no sender identity to publish as. Set coordinator.selfId to a name that "
+                    + "is unique across every daemon sharing {} and restart bridged.",
+                    stripCredentials(uri));
+            return null;
+        }
+        try {
+            LeadMailbox mailbox = opener.open(uri, coordinator.selfId(), coordinator.prefetchOrDefault());
+            log.info("lead coordination: ON as coord-id {} (prefetch={})",
+                    coordinator.selfId(), coordinator.prefetchOrDefault());
+            return mailbox;
+        } catch (IllegalStateException e) {
+            log.warn("cannot reach the AMQP coordination broker ({}) — lead-to-lead messaging is OFF "
+                            + "for this process lifetime. fleet_send{{coordId}} will report it as "
+                            + "not configured, and peer messages already queued stay on the broker until "
+                            + "a restart picks them up. Reason: {}",
+                    stripCredentials(uri), reasonOf(e));
+            return null;
+        }
+    }
+
     /**
      * CB-151/152: pick the reply inbox. A usable broker — a literal {@code uri}, or a {@code
      * uriEnv} whose variable resolves (both read from {@code env}) — selects the durable AMQP inbox.

bridged/src/main/java/dev/ltms/fleet/config/FleetConfig.java

@@ -294,6 +294,19 @@ public record FleetConfig(
      *                       profile that does not opt in. Read live off the current config, so it is
      *                       HOT: a change takes effect on the next exhaustion classification / spawn,
      *                       no restart needed.
+     * @param autoCompactWindow  opt-in per-profile token window that forces a spawned member to
+     *                       auto-compact its context at (Claude Code) or within (opencode) a bound the
+     *                       operator chooses, instead of the backend's own default. {@code null} (the
+     *                       default) leaves today's behaviour exactly — opencode already forces
+     *                       {@code compaction.auto: true} unconditionally (CB-523) but has no absolute
+     *                       window, and Claude Code has neither. When set, validated at config load to
+     *                       {@code [100000, 1000000]} — the band Claude Code's own {@code --autocompact
+     *                       <tokens>} flag accepts. The two backends honour it differently: Claude Code
+     *                       compacts AT this window (a launch-time {@code --autocompact} flag);
+     *                       opencode has no such knob, so this is applied as the model's
+     *                       {@code limit.context} instead, which bounds the window opencode compacts
+     *                       <em>within</em>, and only when {@code model} resolves to a
+     *                       {@code provider/model} pair.
      */
     @JsonIgnoreProperties(ignoreUnknown = true)
     public record Profile(String profile, String baseUrl, String model,
@@ -311,7 +324,8 @@ public record FleetConfig(
                          String credentialId,
                          String ideMcpUrl,
                          String ideProjectDir,
-                         String ideOpenCommand) {
+                         String ideOpenCommand,
+                         Integer autoCompactWindow) {
 
         /** Peer kind spawned by {@link dev.ltms.fleet.member.ClaudeCodeLauncher} (the default). */
         public static final String KIND_CLAUDE_CODE = "claude-code";
@@ -381,6 +395,10 @@ public record FleetConfig(
             // substituted; blank ⇒ no auto-open (the operator opens the module by hand).
             ideProjectDir = (ideProjectDir == null || ideProjectDir.isBlank()) ? null : ideProjectDir;
             ideOpenCommand = (ideOpenCommand == null || ideOpenCommand.isBlank()) ? null : ideOpenCommand;
+            // Opt-in per profile, default off (null). No clamping here — unlike ideMcpUrl/ideProjectDir
+            // there is no blank-string form to normalize (it's an Integer), and the [100000, 1000000]
+            // range is enforced eagerly at config load (rejectAutoCompactWindowOutOfRange), naming the
+            // profile, rather than silently clamped here. A profile that never sets it keeps null.
         }
 
         /**
@@ -425,7 +443,29 @@ public record FleetConfig(
         public Profile withProfile(String p) {
             return new Profile(p, baseUrl, model, configDir, tokenEnv, argv, placement, workspace, tabLabel,
                     mcpUrl, cwd, parityOverlay, gitTokenEnv, gitHostEnv, kind, env, weight, maxLoad, subscription,
-                    exhaustedPattern, credentialId, ideMcpUrl, ideProjectDir, ideOpenCommand);
+                    exhaustedPattern, credentialId, ideMcpUrl, ideProjectDir, ideOpenCommand, autoCompactWindow);
+        }
+
+        /**
+         * Backward-compatible constructor without the {@code autoCompactWindow} field — the profile
+         * leaves auto-compaction at the backend's own default (opencode's unconditional
+         * {@code compaction.auto: true}, or Claude Code's built-in threshold), exactly as before this
+         * key existed. This is the shape the canonical constructor had before the field was added —
+         * every pre-existing Java call site (and any YAML that omits the key) keeps compiling and
+         * behaving identically; Jackson binds the canonical (longest) constructor, so YAML omitting
+         * {@code autoCompactWindow:} still lands here as {@code null} via that path, not this one.
+         */
+        public Profile(String profile, String baseUrl, String model,
+                      String configDir, String tokenEnv, List<String> argv,
+                      String placement, String workspace, String tabLabel, String mcpUrl,
+                      String cwd, List<String> parityOverlay, String gitTokenEnv, String gitHostEnv,
+                      String kind, Map<String, String> env, Float weight, Integer maxLoad,
+                      Boolean subscription, String exhaustedPattern, String credentialId,
+                      String ideMcpUrl, String ideProjectDir, String ideOpenCommand) {
+            this(profile, baseUrl, model, configDir, tokenEnv, argv, placement, workspace, tabLabel,
+                    mcpUrl, cwd, parityOverlay, gitTokenEnv, gitHostEnv, kind, env, weight, maxLoad,
+                    subscription, exhaustedPattern, credentialId, ideMcpUrl, ideProjectDir, ideOpenCommand,
+                    null);
         }
 
         /**
@@ -1290,6 +1330,7 @@ public record FleetConfig(
             warnUnknownTopLevelKeys(yaml, path);
             rejectDuplicateMemberSlots(yaml);
             rejectNegativeMaxLoad(yaml);
+            rejectAutoCompactWindowOutOfRange(yaml);
             rejectUnknownKind(yaml);
             rejectUnknownAuthMode(yaml);
             rejectUnknownPlacement(yaml);
@@ -1592,6 +1633,52 @@ public record FleetConfig(
         }
     }
 
+    /** Lowest {@code autoCompactWindow} Claude Code's {@code --autocompact <tokens>} flag accepts. */
+    static final int AUTO_COMPACT_WINDOW_MIN = 100_000;
+    /** Highest {@code autoCompactWindow} Claude Code's {@code --autocompact <tokens>} flag accepts. */
+    static final int AUTO_COMPACT_WINDOW_MAX = 1_000_000;
+
+    /**
+     * Reject a profile whose {@code autoCompactWindow:} is set but outside the token band Claude
+     * Code's own {@code --autocompact <tokens>} flag accepts (100k–1M), naming both the profile and
+     * the value.
+     *
+     * <p>Unset/{@code null} means "off" and passes silently — today's behaviour for every profile
+     * that does not opt in (see {@link Profile#autoCompactWindow()}). A profile that DOES set the key
+     * is validated eagerly, at config load, rather than failing later when Claude Code itself refuses
+     * the launch flag on spawn — the same "fail loud at load, not lazily at first spawn" reasoning as
+     * {@link #rejectNegativeMaxLoad} and {@link #rejectUnknownPlacementPolicy}.
+     *
+     * @param yaml the raw config text
+     * @throws IllegalStateException when any profile's {@code autoCompactWindow} is set and outside
+     *                                {@code [100000, 1000000]}
+     */
+    static void rejectAutoCompactWindowOutOfRange(String yaml) {
+        Map<?, ?> raw;
+        try {
+            raw = YAML.readValue(yaml, Map.class);
+        } catch (IOException | IllegalArgumentException e) {
+            return; // a malformed file is reported by the real parse, not here
+        }
+        if (raw == null || !(raw.get("profiles") instanceof Map<?, ?> profiles)) {
+            return;
+        }
+        List<String> bad = profiles.entrySet().stream()
+                .filter(e -> e.getValue() instanceof Map<?, ?> p
+                        && p.get("autoCompactWindow") instanceof Number n
+                        && (n.doubleValue() < AUTO_COMPACT_WINDOW_MIN || n.doubleValue() > AUTO_COMPACT_WINDOW_MAX))
+                .map(e -> String.valueOf(e.getKey()))
+                .sorted()
+                .toList();
+        if (!bad.isEmpty()) {
+            throw new IllegalStateException("refusing to start: profile(s) [" + String.join(", ", bad)
+                    + "] set autoCompactWindow outside [" + AUTO_COMPACT_WINDOW_MIN + ", "
+                    + AUTO_COMPACT_WINDOW_MAX + "] — Claude Code's --autocompact flag accepts only "
+                    + "that band of tokens; omit the key to leave auto-compaction at each backend's "
+                    + "own default.");
+        }
+    }
+
     /** The peer kinds this build has an adapter for — {@link Profile#kind()}'s only valid values. */
     private static final Set<String> KNOWN_KINDS = Set.of(Profile.KIND_CLAUDE_CODE, Profile.KIND_OPENCODE);
 

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

@@ -11,6 +11,8 @@ import dev.ltms.fleet.metrics.FleetMetrics;
 import dev.ltms.fleet.metrics.Metrics;
 import dev.ltms.fleet.inject.MemberPresence;
 import dev.ltms.fleet.herdr.HerdrException;
+import dev.ltms.fleet.msg.LeadChannel;
+import dev.ltms.fleet.msg.LeadMessage;
 import dev.ltms.fleet.msg.MessageService;
 import dev.ltms.fleet.msg.Rendezvous;
 import dev.ltms.fleet.peer.PeerUnreachableException;
@@ -37,6 +39,7 @@ import java.util.List;
 import java.util.Map;
 import java.util.Objects;
 import java.util.Set;
+import java.util.UUID;
 import java.util.concurrent.ConcurrentHashMap;
 import java.util.function.BiFunction;
 import java.util.function.Function;
@@ -97,6 +100,8 @@ public final class FleetMcp {
     private final CapacitySource capacity;
     private final HealthCoverageSource healthCoverage;
     private final QuarantineSource quarantine;
+    /** CB-637: this daemon's lead-to-lead channel; {@code null} when no coordinator is configured. */
+    private final LeadChannel leadChannel;
 
     /** Capacity facts used by {@code fleet_list}; production must supply the placement live count. */
     public record CapacitySource(Function<String, Integer> liveCount, Function<String, Integer> maxLoad,
@@ -131,6 +136,21 @@ public final class FleetMcp {
                      ConnectionIdentity identity, MemberPresence presence, PrimaryRegistry primaryRegistry,
                       CallerResolver callers, Metrics metrics, CapacitySource capacity, HealthCoverageSource healthCoverage,
                       QuarantineSource quarantine) {
+        this(messages, workers, sessions, identity, presence, primaryRegistry, callers, metrics, capacity,
+                healthCoverage, quarantine, null);
+    }
+
+    /**
+     * As above, with this daemon's lead-to-lead channel (CB-637). {@code leadChannel} is
+     * {@code null} whenever no {@code coordinator:} block is configured or its broker could not be
+     * reached at boot — cross-daemon lead messaging is simply off, and {@code fleet_send{coordId}}
+     * says so rather than failing obscurely.
+     */
+    public FleetMcp(MessageService messages, PeerLauncher workers, SessionManager sessions,
+                     ConnectionIdentity identity, MemberPresence presence, PrimaryRegistry primaryRegistry,
+                      CallerResolver callers, Metrics metrics, CapacitySource capacity, HealthCoverageSource healthCoverage,
+                      QuarantineSource quarantine, LeadChannel leadChannel) {
+        this.leadChannel = leadChannel;
         this.capacity = capacity;
         this.quarantine = Objects.requireNonNull(quarantine, "quarantine");
         this.healthCoverage = healthCoverage;
@@ -177,6 +197,13 @@ public final class FleetMcp {
                     String target = str(a, "sessionId");
                     String content = str(a, "content");
                     String turnId = str(a, "turnId");
+                    String coordId = str(a, "coordId");
+                    if (coordId != null && !coordId.isBlank()) {
+                        // CB-637: a peer LEAD on another daemon, addressed by coord-id over the shared
+                        // coordination broker. Checked before the turnId branch so a call that sets both
+                        // is rejected as the conflict it is, rather than silently taking one route.
+                        return sendToLead(leadChannel, coordId, content, target, turnId);
+                    }
                     if (turnId != null && !turnId.isBlank()) {
                         // Answering a worker's fleet_ask (CB-205): resolve its blocked question and
                         // block for the worker's reply as it resumes the same turn. This is the same
@@ -257,7 +284,8 @@ public final class FleetMcp {
                     if (denied != null) return denied;
                     return listFleet(workers, sessions, messages, capacity, healthCoverage, quarantine,
                             callers == null ? Map.of() : callers.leads(),
-                            callerTerminal(exchange));
+                            callerTerminal(exchange),
+                            leadChannel == null ? null : leadChannel.selfCoordId());
                 };
         BiFunction<McpSyncServerExchange, McpSchema.CallToolRequest, McpSchema.CallToolResult> stopHandler =
                 (exchange, req) -> {
@@ -585,6 +613,54 @@ public final class FleetMcp {
         return null;
     }
 
+    /**
+     * {@code fleet_send} carrying a {@code coordId} (CB-637): a message to a PEER LEAD, published to
+     * that lead's durable mailbox on the shared coordination broker. This is the only lead→lead path
+     * that crosses hosts — the existing pane-injection route can only reach a lead whose herdr socket
+     * this daemon shares.
+     *
+     * <p>{@code coordId} is mutually exclusive with {@code sessionId} and {@code turnId}: those two
+     * address a worker session owned by <em>this</em> daemon, a coord-id addresses a lead owned by
+     * another one, and there is no sensible reading of a call that sets both. Rejected by name rather
+     * than resolved by precedence, so a caller that meant the other route learns it instead of having
+     * its message quietly go somewhere else.
+     *
+     * <p>The publish is synchronous and confirmed by the broker, so the result is a real delivery
+     * receipt rather than a hopeful one. Its failure — nobody owns {@code coordId}'s mailbox, the
+     * broker nacked, or the confirm timed out — arrives as {@link IllegalStateException} and is
+     * turned into a tool error naming the coord-id. It is never allowed to escape as a crash: an
+     * unreachable peer is an ordinary outcome of addressing a fleet you do not control.
+     *
+     * @param leadChannel this daemon's channel, or {@code null} when no coordinator is configured
+     */
+    static McpSchema.CallToolResult sendToLead(LeadChannel leadChannel, String coordId, String content,
+                                               String sessionId, String turnId) {
+        if (!isBlank(sessionId) || !isBlank(turnId)) {
+            String conflict = !isBlank(sessionId) ? "sessionId" : "turnId";
+            return error("coordId and " + conflict + " are mutually exclusive: coordId addresses a peer "
+                    + "LEAD on another daemon over the coordination broker, while " + conflict
+                    + " addresses a worker session on this one. Pass exactly one.");
+        }
+        if (isBlank(content)) {
+            return error("content is required");
+        }
+        if (leadChannel == null) {
+            return error("lead coordination is not configured (no coordinator: block) — cannot send to "
+                    + "peer lead \"" + coordId + "\". Add a coordinator: block with a shared broker uri "
+                    + "and this daemon's selfId, then restart bridged.");
+        }
+        LeadMessage msg = new LeadMessage(UUID.randomUUID().toString(), leadChannel.selfCoordId(),
+                coordId, content);
+        try {
+            leadChannel.publish(coordId, msg);
+        } catch (IllegalStateException e) {
+            return error("cannot deliver to peer lead \"" + coordId + "\": " + e.getMessage()
+                    + ". Check that a daemon is running with coordinator.selfId=\"" + coordId
+                    + "\" and is connected to the same coordination broker.");
+        }
+        return text("delivered to peer lead " + coordId + " (msgId " + msg.msgId() + ")");
+    }
+
     /** {@code fleet_poll}: check an async delegation by ticket, or drain a worker's inbox by target. */
     static McpSchema.CallToolResult poll(MessageService messages, String ticket, String target) {
         if (!isBlank(target)) {
@@ -870,6 +946,22 @@ public final class FleetMcp {
     static McpSchema.CallToolResult listFleet(PeerLauncher workers, SessionManager sessions, MessageService messages,
                                                CapacitySource capacity, HealthCoverageSource healthCoverage,
                                                QuarantineSource quarantine, Map<String, String> leads, String selfTerm) {
+        return listFleet(workers, sessions, messages, capacity, healthCoverage, quarantine, leads, selfTerm, null);
+    }
+
+    /**
+     * As above, additionally reporting this daemon's own lead coordination id (CB-637) when one is
+     * configured and its channel opened. There is no peer-discovery surface yet — a lead addresses a
+     * peer by a coord-id it was told — so this row exists to answer the one question the operator
+     * cannot answer any other way: what is MY coord-id, the one a peer must use to reach me. It is
+     * omitted entirely when no coordinator is configured, so an ordinary fleet's output is unchanged.
+     *
+     * @param selfCoordId this daemon's coord-id, or {@code null} when lead coordination is off
+     */
+    static McpSchema.CallToolResult listFleet(PeerLauncher workers, SessionManager sessions, MessageService messages,
+                                               CapacitySource capacity, HealthCoverageSource healthCoverage,
+                                               QuarantineSource quarantine, Map<String, String> leads, String selfTerm,
+                                               String selfCoordId) {
         try {
             Map<String, Agent> live = workers.list().stream()
                     .map(Agent.class::cast)
@@ -888,6 +980,9 @@ public final class FleetMcp {
             Map<String, Object> result = new LinkedHashMap<>();
             result.put("leads", leadRows); result.put("members", out);
             result.put("healthCoverage", healthCoverage.value().get());
+            if (selfCoordId != null && !selfCoordId.isBlank()) {
+                result.put("coordinator", Map.of("selfId", selfCoordId, "configured", true));
+            }
             if (capacity.available()) result.put("capacity", profiles.stream()
                     .map(profile -> capacityView(profile, capacity.liveCount(), capacity.maxLoad(), roster, messages,
                             capacity.clock().getAsLong(), quarantine)).toList());
@@ -1015,7 +1110,9 @@ public final class FleetMcp {
                 "Delegate a task to a worker session. By default blocks until the worker replies and "
                         + "returns its reply (or a 'still working / queued' note on timeout). Pass wait:false "
                         + "for a long task to return a ticket immediately, then poll it with fleet_poll. To "
-                        + "answer a worker's fleet_ask, pass its turnId (with content) instead of sessionId.",
+                        + "answer a worker's fleet_ask, pass its turnId (with content) instead of sessionId. "
+                        + "To message a PEER LEAD on another daemon — possibly another host — pass its "
+                        + "coordId instead; that is coordination, never a task.",
                 objectSchema(Map.of(
                         "sessionId", stringProp("The worker session id (herdr terminal_id) to delegate to"),
                         "content", stringProp("The task/message to send to the worker (or your answer, with turnId)"),
@@ -1023,7 +1120,11 @@ public final class FleetMcp {
                         "wait", Map.of("type", "boolean",
                                 "description", "Block for the reply (default true); false returns a ticket to poll"),
                         "turnId", stringProp("When answering a worker's fleet_ask, its question turnId — "
-                                + "routes your answer back into the same turn (omit for a normal delegation)")),
+                                + "routes your answer back into the same turn (omit for a normal delegation)"),
+                        "coordId", stringProp("A peer LEAD's coordination id — delivers content to that "
+                                + "lead's durable mailbox on the shared coordination broker, which works "
+                                + "across hosts. Mutually exclusive with sessionId and turnId. Your own "
+                                + "coordId is reported by fleet_list.")),
                         List.of("content")));
     }
 

bridged/src/main/java/dev/ltms/fleet/member/ClaudeCodeLauncher.java

@@ -248,7 +248,7 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
         // has neither MCP nor a charter — session flags must be added into a list we own.
         List<String> argv = mutableArgv(argvWithFleet(cfg, spec));
         String agentSessionId = applySessionIdentity(argv, spec.sessionName(), spec.resumeSessionId());
-        return new Launch(workerEnv, argvWithModel(argv, cfg), agentSessionId);
+        return new Launch(workerEnv, argvWithAutoCompact(argvWithModel(argv, cfg), cfg), agentSessionId);
     }
 
     /**
@@ -463,6 +463,30 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
         return withModel;
     }
 
+    /**
+     * Pin a bounded auto-compaction window on the command line via {@code --autocompact <tokens>},
+     * opt-in per profile (CB-634's sibling ticket: a member that runs out of context dies mid-turn
+     * and its {@code fleet_reply} — the whole point of the turn — is lost with it; opencode already
+     * forces {@code compaction.auto: true} unconditionally, CB-523, but Claude Code has no equivalent
+     * and runs at the backend's own default window).
+     *
+     * <p>Mirrors {@link #argvWithModel}: appended after it, so it survives the {@code ccs <profile>}
+     * wrapper the same way {@code --model} does, and outranks env/settings and the operator's own
+     * {@code argv}. Verified: {@code claude 2.1.241 --help} lists {@code --autocompact <auto|tokens>}
+     * (either the literal {@code auto}, or an integer 100k–1M) — {@link FleetConfig#load} rejects a
+     * configured value outside that band before this ever runs, so the flag Claude Code receives here
+     * is always in range.
+     */
+    private static List<String> argvWithAutoCompact(List<String> argv, FleetConfig.Profile cfg) {
+        if (cfg.autoCompactWindow() == null) {
+            return argv;
+        }
+        List<String> withAutoCompact = mutableArgv(argv);
+        withAutoCompact.add("--autocompact");
+        withAutoCompact.add(String.valueOf(cfg.autoCompactWindow()));
+        return withAutoCompact;
+    }
+
     // --- Agent-returning convenience spawns (used by callers/tests that want the herdr Agent) ---
 
     /** Spawn a worker for the default profile in the resolved default cwd. */

bridged/src/main/java/dev/ltms/fleet/member/OpenCodeLauncher.java

@@ -209,9 +209,19 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
     @Override
     protected Launch buildLaunch(FleetConfig.Profile cfg, LaunchSpec spec) {
         Map<String, String> workerEnv = baseEnv(cfg);
+        // autoCompactWindow's opencode lever (limit.context) only targets a specific provider/model
+        // entry, so it needs model: in "provider/model" form. A profile that opts in without that
+        // shape gets no silent no-op — log it, once, here, whether or not writeConfig ends up running.
+        boolean wantsContextLimit = cfg.autoCompactWindow() != null && splitProviderModel(cfg.model()) != null;
+        if (cfg.autoCompactWindow() != null && !wantsContextLimit) {
+            log.warn("profile '{}' sets autoCompactWindow but model '{}' is not \"<provider>/<model>\" "
+                            + "form — opencode's per-model context limit could not be applied for this profile",
+                    cfg.profile(), cfg.model());
+        }
         // A config file is needed for the bridge MCP mount, a member charter, the IDE MCP (+ its
-        // guidance overlay, CB-634), or a pinned endpoint (CB-508).
-        if (cfg.hasMcp() || cfg.hasIdeMcp() || spec.charter() != null || hasCustomProvider(cfg)) {
+        // guidance overlay, CB-634), a pinned endpoint (CB-508), or a resolvable autoCompactWindow.
+        if (cfg.hasMcp() || cfg.hasIdeMcp() || spec.charter() != null || hasCustomProvider(cfg)
+                || wantsContextLimit) {
             workerEnv.put("OPENCODE_CONFIG", writeConfig(cfg, spec.charter(), spec.cwd()).toString());
         }
         applyGitToken(workerEnv, cfg);
@@ -360,6 +370,9 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
             if (hasCustomProvider(cfg)) {
                 addCustomProvider(root, cfg);
             }
+            if (cfg.autoCompactWindow() != null) {
+                applyContextLimit(root, cfg);
+            }
 
             Path cfgFile = dir.resolve("opencode.json");
             // Built with Jackson rather than string concatenation: the provider block is nested and
@@ -404,18 +417,68 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
      * default gateway — a worker quietly talking to the wrong endpoint is the failure this avoids.
      */
     private static String[] splitModelSelector(FleetConfig.Profile cfg) {
-        String model = cfg.model();
-        int slash = model == null ? -1 : model.indexOf('/');
-        if (model == null || model.isBlank() || slash <= 0 || slash == model.length() - 1) {
+        String[] parts = splitProviderModel(cfg.model());
+        if (parts == null) {
             throw new IllegalArgumentException(
                     "profile " + cfg.profile() + " sets baseUrl (a pinned opencode endpoint) so"
                             + " model: must be \"<provider>/<model>\", e.g."
                             + " \"local-vllm/deepseek-v4-flash\"; got "
-                            + (model == null ? "null" : '"' + model + '"'));
+                            + (cfg.model() == null ? "null" : '"' + cfg.model() + '"'));
+        }
+        return parts;
+    }
+
+    /**
+     * Split {@code model} into its {@code provider} and {@code model} halves, or {@code null} when
+     * it is not in that shape (unset/blank, or no non-trailing {@code /}). Unlike
+     * {@link #splitModelSelector}, non-throwing — callers that only *optionally* need the split
+     * (autoCompactWindow's context-limit application) use this to fall back to a WARN rather than an
+     * exception, since a profile without {@code baseUrl} is not required to name a provider/model.
+     */
+    private static String[] splitProviderModel(String model) {
+        int slash = model == null ? -1 : model.indexOf('/');
+        if (model == null || model.isBlank() || slash <= 0 || slash == model.length() - 1) {
+            return null;
         }
         return new String[]{model.substring(0, slash), model.substring(slash + 1)};
     }
 
+    /**
+     * Apply the per-profile {@code autoCompactWindow} as opencode's per-model context limit.
+     *
+     * <p>opencode has no absolute "compact at N tokens" knob — its {@code compaction} block only
+     * exposes {@code auto}/{@code prune}/{@code reserved}/{@code tail_turns}/
+     * {@code preserve_recent_tokens} — so the real lever is the model's own
+     * {@code provider.<p>.models.<m>.limit.context}, which bounds the window opencode compacts
+     * <em>within</em> rather than compacting exactly AT it the way Claude Code's {@code --autocompact}
+     * does.
+     *
+     * <p>Uses get-or-create nodes ({@code withObject}) at every level so this MERGES with any provider
+     * block {@link #addCustomProvider} already wrote for a custom-provider (pinned-endpoint) profile —
+     * it must never overwrite that block's {@code npm}/{@code name}/{@code options}. For a gateway
+     * profile (no {@code baseUrl}, so no prior provider block) this writes a partial
+     * {@code provider.<p>.models.<m>.limit} override, which opencode merges over its own built-in
+     * provider definition.
+     *
+     * <p>opencode's {@code limit} schema requires both {@code context} and {@code output}; there is no
+     * independent signal for the latter here, so 16384 is written as a safe default (documented in
+     * {@code fleetd.example.yaml}).
+     *
+     * <p>Silently does nothing when {@code model:} is not in {@code provider/model} form — a warning
+     * for that case is already logged once in {@code buildLaunch}, so this stays quiet rather than
+     * duplicating it.
+     */
+    private static void applyContextLimit(ObjectNode root, FleetConfig.Profile cfg) {
+        String[] parts = splitProviderModel(cfg.model());
+        if (parts == null) {
+            return;
+        }
+        ObjectNode limit = root.withObject("provider").withObject(parts[0])
+                .withObject("models").withObject(parts[1]).withObject("limit");
+        limit.put("context", cfg.autoCompactWindow());
+        limit.put("output", 16384);
+    }
+
     /**
      * The OpenAI-compatible base URL for {@code baseUrl}. A bare {@code host:port} gets {@code /v1}
      * appended (where these servers put the API); a URL that already carries a path is taken as-is,

bridged/src/main/java/dev/ltms/fleet/msg/LeadChannel.java

@@ -0,0 +1,40 @@
+package dev.ltms.fleet.msg;
+
+import java.util.List;
+
+/**
+ * The lead-to-lead message channel this daemon speaks, as its callers need it — one lead's own
+ * mailbox: publish to a peer's coord-id, look at what has arrived for me, and ack what I have
+ * delivered.
+ *
+ * <p>Extracted from {@link LeadMailbox} purely as a seam. {@code LeadMailbox} is the one production
+ * implementation and owns a live AMQP connection, so a test that wanted to exercise the routing in
+ * {@code FleetMcp} or the delivery in {@link LeadCoordLoop} would have had to stand up a broker —
+ * which is exactly the kind of test that gets tagged {@code contract} and then does not run. With
+ * this interface both of those are hermetic: they inject a fake channel and assert on what was
+ * published, peeked and acked.
+ *
+ * <p>Note what is <em>not</em> here: {@code drain()} and {@code close()}. Draining is a convenience
+ * over peek+ack that no caller on this seam uses, and closing is the owner's job — {@code Fleetd}
+ * holds the concrete {@link LeadMailbox} for its shutdown hook and hands only this narrower view to
+ * everyone else.
+ */
+public interface LeadChannel {
+
+    /**
+     * Send {@code m} to {@code toCoordId}'s mailbox, blocking until the broker confirms it is
+     * durably queued. Throws {@link IllegalStateException} when it is not — unroutable (nobody owns
+     * that coord-id), nacked, or unconfirmed within the implementation's timeout. A caller must
+     * report that as a failed send, never as a delivered one.
+     */
+    void publish(String toCoordId, LeadMessage m);
+
+    /** Non-destructive FIFO snapshot of the messages held for this daemon's own coord-id. */
+    List<LeadMessage> peek();
+
+    /** Drop {@code msgId} from the held set and ack it on the broker. A no-op if it is not held. */
+    void ack(String msgId);
+
+    /** This daemon's own lead coordination id — the mailbox it owns, and the {@code from} it sends as. */
+    String selfCoordId();
+}

bridged/src/main/java/dev/ltms/fleet/msg/LeadCoordLoop.java

@@ -0,0 +1,193 @@
+package dev.ltms.fleet.msg;
+
+import dev.ltms.fleet.herdr.AgentControl;
+import dev.ltms.fleet.herdr.AgentStatus;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.concurrent.TimeUnit;
+import java.util.function.Supplier;
+
+/**
+ * The receive half of lead-to-lead messaging: a bounded background loop that takes what has arrived
+ * in this daemon's own {@link LeadChannel} mailbox and types it into the local lead's herdr pane.
+ *
+ * <p>{@code fleet_send{coordId}} is the send half — it publishes to a peer daemon's mailbox and
+ * returns. Nothing on the receiving side reads that mailbox on its own, because the peer lead is an
+ * interactive agent, not a service that polls; this loop is what closes the gap.
+ *
+ * <p><strong>Status-gated, exactly like {@link ReplyPushLoop}.</strong> A pane may only be injected
+ * into at a turn boundary ({@link AgentStatus#injectable()} — idle, blocked or done); pasting into
+ * a live turn corrupts it. So a tick that finds the lead busy simply does nothing and comes back
+ * later.
+ *
+ * <p><strong>Ack only after delivery.</strong> A message is acked — removed from the broker — only
+ * once {@link AgentControl#send} has actually put it in the pane. Anything not delivered (no lead
+ * pane resolvable, lead mid-turn, herdr threw) stays unacked and is retried on the next tick, and
+ * survives a daemon restart because the broker still holds it. The cost of that choice is a
+ * possible duplicate — the send lands and the ack does not — which is the right way round: a peer
+ * lead seeing a message twice is a nuisance, a peer lead never seeing it at all is the failure this
+ * whole path exists to remove.
+ *
+ * <p><strong>One message per tick.</strong> The loop delivers at most one held message per tick even
+ * when several are waiting. Injecting a second one immediately would mean acting on a status read
+ * taken <em>before</em> the first injection: that first paste starts a turn, and herdr does not
+ * report the pane as {@code working} the instant it does. Waiting for the next tick means every
+ * delivery is gated on a status read that already saw the previous one. A backlog therefore drains
+ * one message per {@code intervalMs}, in FIFO order.
+ */
+public final class LeadCoordLoop {
+
+    private static final Logger log = LoggerFactory.getLogger(LeadCoordLoop.class);
+
+    /** How an arriving peer message is rendered into the lead's pane — the sender's coord-id, then its text. */
+    static final String DELIVERY_FORMAT = "[lead %s] %s";
+
+    private final LeadChannel channel;
+    private final AgentControl agents;
+    private final Supplier<Map<String, String>> leads;
+    private final ScheduledExecutorService scheduler;
+    private final long intervalMs;
+
+    private volatile boolean running;
+
+    /**
+     * @param channel  this daemon's own lead mailbox
+     * @param agents   herdr control, for the status gate and the pane injection
+     * @param leads    live {@code terminal_id → name} view of the leads this daemon recognises —
+     *                 read through the supplier on every tick, never snapshotted, so a lead found by
+     *                 the tab scan after startup becomes reachable without a restart
+     * @param scheduler the loop's own scheduler; the caller owns its shutdown
+     * @param intervalMs how long between ticks
+     */
+    public LeadCoordLoop(LeadChannel channel, AgentControl agents, Supplier<Map<String, String>> leads,
+                         ScheduledExecutorService scheduler, long intervalMs) {
+        this.channel = channel;
+        this.agents = agents;
+        this.leads = leads;
+        this.scheduler = scheduler;
+        this.intervalMs = intervalMs;
+    }
+
+    /** Begin ticking. Idempotent-ish: calling it twice would schedule two chains, so call it once. */
+    public void start() {
+        running = true;
+        log.info("lead coordination: delivering peer messages for coord-id {} every {}ms",
+                channel.selfCoordId(), intervalMs);
+        scheduleNext();
+    }
+
+    /** Stop ticking. In-flight work finishes; nothing further is scheduled. */
+    public void close() {
+        running = false;
+    }
+
+    private void scheduleNext() {
+        if (!running) {
+            return;
+        }
+        scheduler.schedule(this::tickAndReschedule, intervalMs, TimeUnit.MILLISECONDS);
+    }
+
+    private void tickAndReschedule() {
+        try {
+            tick();
+        } catch (RuntimeException e) {
+            // Never let one bad tick end the chain — the next one re-reads everything from scratch.
+            log.warn("lead coordination tick failed: {}", e.toString());
+        }
+        scheduleNext();
+    }
+
+    /**
+     * One tick: deliver at most one held peer message into the local lead's pane and ack it.
+     * Package-private so a test drives it directly rather than waiting on the scheduler.
+     */
+    void tick() {
+        List<LeadMessage> held;
+        try {
+            held = channel.peek();
+        } catch (RuntimeException e) {
+            log.debug("lead coordination: cannot read the mailbox this tick: {}", e.toString());
+            return;
+        }
+        if (held.isEmpty()) {
+            return;
+        }
+        String lead = resolveLocalLead();
+        if (lead == null) {
+            // Left unacked on purpose: the broker keeps holding it until a lead pane exists.
+            log.debug("lead coordination: {} message(s) waiting but no local lead pane to deliver to",
+                    held.size());
+            return;
+        }
+        AgentStatus status;
+        try {
+            status = agents.status(lead);
+        } catch (RuntimeException e) {
+            log.debug("lead coordination: status check failed for lead {}, retrying next tick", lead, e);
+            return;
+        }
+        if (!status.injectable()) {
+            log.debug("lead coordination: lead {} is {} (not injectable), holding {} message(s)",
+                    lead, status, held.size());
+            return;
+        }
+        LeadMessage msg = held.getFirst();
+        try {
+            agents.send(lead, DELIVERY_FORMAT.formatted(msg.from(), msg.content()));
+        } catch (RuntimeException e) {
+            // Not delivered, so not acked — the broker still has it for the next tick.
+            log.warn("lead coordination: failed to deliver message {} from {} to lead {}: {}",
+                    msg.msgId(), msg.from(), lead, e.toString());
+            return;
+        }
+        try {
+            channel.ack(msg.msgId());
+        } catch (RuntimeException e) {
+            // Delivered but not acked: it will be redelivered, which the javadoc calls out as the
+            // deliberate direction of this trade.
+            log.warn("lead coordination: delivered message {} but could not ack it: {}",
+                    msg.msgId(), e.toString());
+            return;
+        }
+        log.debug("lead coordination: delivered message {} from {} to lead {}", msg.msgId(), msg.from(), lead);
+    }
+
+    /**
+     * Which local pane a peer's message is for. The mailbox's {@code selfCoordId} is this daemon's
+     * one lead identity, so there is exactly one right answer — this only has to find it:
+     *
+     * <ol>
+     *   <li>a lead whose configured name equals {@code selfCoordId} — the explicit, unambiguous case;</li>
+     *   <li>otherwise the sole lead, when this daemon recognises exactly one;</li>
+     *   <li>otherwise nothing, and the message waits.</li>
+     * </ol>
+     *
+     * <p>Step 3 is deliberate rather than a guess-the-lead fallback. Picking one of several leads
+     * arbitrarily would type a peer's message into a pane it was not addressed to, and the message
+     * would then be acked and gone. Leaving it held costs a delay and nothing else.
+     */
+    private String resolveLocalLead() {
+        Map<String, String> known = leads.get();
+        if (known.isEmpty()) {
+            return null;
+        }
+        String self = channel.selfCoordId();
+        for (var entry : known.entrySet()) {
+            if (entry.getValue() != null && entry.getValue().equals(self)) {
+                return entry.getKey();
+            }
+        }
+        if (known.size() == 1) {
+            return known.keySet().iterator().next();
+        }
+        log.warn("lead coordination: {} leads are known and none is named \"{}\" — cannot tell which "
+                        + "pane a peer message is for; name one lead after coordinator.selfId to fix this",
+                known.size(), self);
+        return null;
+    }
+}

bridged/src/main/java/dev/ltms/fleet/msg/LeadMailbox.java

@@ -61,7 +61,7 @@ import java.util.concurrent.TimeoutException;
  * redelivery) and any publish still awaiting its confirm is failed rather than left to idle out
  * the confirm timeout against a sequence number that means nothing on the new channel.
  */
-public final class LeadMailbox implements AutoCloseable {
+public final class LeadMailbox implements LeadChannel, AutoCloseable {
 
     private static final Logger log = LoggerFactory.getLogger(LeadMailbox.class);
 
@@ -195,6 +195,7 @@ public final class LeadMailbox implements AutoCloseable {
      * confirmed within {@link #CONFIRM_TIMEOUT_MS} — the caller must treat that as a failed publish,
      * not a lost-and-forgotten one.
      */
+    @Override
     public void publish(String toCoordId, LeadMessage msg) {
         byte[] body;
         try {
@@ -239,7 +240,14 @@ public final class LeadMailbox implements AutoCloseable {
         }
     }
 
+    /** The coord-id whose mailbox this instance owns — the {@code from} of everything it publishes. */
+    @Override
+    public String selfCoordId() {
+        return selfCoordId;
+    }
+
     /** Non-destructive FIFO snapshot of this mailbox's currently-held messages. */
+    @Override
     public List<LeadMessage> peek() {
         synchronized (held) {
             return held.values().stream().map(Held::message).toList();
@@ -254,6 +262,7 @@ public final class LeadMailbox implements AutoCloseable {
     }
 
     /** Remove the held message {@code msgId} and ack it on the broker. No-op if not held. */
+    @Override
     public void ack(String msgId) {
         Held h;
         synchronized (held) {

bridged/src/test/java/dev/ltms/fleet/FleetdLeadMailboxSelectionTest.java

@@ -0,0 +1,144 @@
+package dev.ltms.fleet;
+
+import ch.qos.logback.classic.Level;
+import ch.qos.logback.classic.Logger;
+import ch.qos.logback.classic.spi.ILoggingEvent;
+import ch.qos.logback.core.read.ListAppender;
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.msg.LeadMailbox;
+import org.junit.jupiter.api.Test;
+import org.slf4j.LoggerFactory;
+
+import java.util.List;
+import java.util.Map;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+/**
+ * CB-637: the daemon decides whether lead-to-lead messaging is on in
+ * {@link Fleetd#openLeadMailbox}, not in the config record — so testing
+ * {@code Coordinator.isConfigured()} alone would pass even if {@code Fleetd} never honoured it.
+ * These drive the real selection with an injected env map and an injected opener, so no broker is
+ * involved and no process environment is mutated.
+ *
+ * <p>The invariant every case shares: the feature turns itself OFF, never takes the daemon down.
+ * A fleet whose coordination broker is missing, half-configured or unreachable must still start and
+ * still work exactly as it did before this feature existed.
+ */
+class FleetdLeadMailboxSelectionTest {
+
+    private static final String SECRET = "c00rdPw";
+    private static final String RESOLVED_URI = "amqp://user:" + SECRET + "@coord.example:5672/coord";
+
+    /** Fake opener: records what it was offered, or fails as an unreachable broker would. */
+    private static final class RecordingOpener implements Fleetd.LeadMailboxOpener {
+        String offeredUri;
+        String offeredSelfId;
+        int offeredPrefetch = -1;
+        boolean unreachable;
+
+        @Override
+        public LeadMailbox open(String uri, String selfCoordId, int prefetch) {
+            this.offeredUri = uri;
+            this.offeredSelfId = selfCoordId;
+            this.offeredPrefetch = prefetch;
+            if (unreachable) {
+                throw new IllegalStateException("cannot connect to AMQP coordination broker at " + uri,
+                        new java.net.ConnectException("Connection refused"));
+            }
+            // A real LeadMailbox needs a live connection; nothing here dereferences the result
+            // beyond a null check, so the "reachable" cases assert on what was OFFERED instead.
+            return null;
+        }
+    }
+
+    private static ListAppender<ILoggingEvent> captureFleetdLogs() {
+        Logger logger = (Logger) LoggerFactory.getLogger(Fleetd.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        return appender;
+    }
+
+    private static String joined(ListAppender<ILoggingEvent> appender, Level level) {
+        return appender.list.stream().filter(e -> e.getLevel() == level)
+                .map(ILoggingEvent::getFormattedMessage).reduce("", (a, b) -> a + "\n" + b);
+    }
+
+    @Test
+    void noCoordinatorBlockLeavesTheFeatureOffSilently() {
+        var appender = captureFleetdLogs();
+        var opener = new RecordingOpener();
+
+        assertNull(Fleetd.openLeadMailbox(null, Map.of(), opener));
+
+        assertNull(opener.offeredUri, "nothing configured means nothing is opened");
+        assertEquals("", joined(appender, Level.WARN),
+                "an opt-in feature nobody asked for must not warn on every boot");
+    }
+
+    @Test
+    void opensTheMailboxWhenAUriAndSelfIdAreConfigured() {
+        var opener = new RecordingOpener();
+        var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, "mac-opus", null);
+
+        Fleetd.openLeadMailbox(coordinator, Map.of(), opener);
+
+        assertEquals(RESOLVED_URI, opener.offeredUri);
+        assertEquals("mac-opus", opener.offeredSelfId);
+        assertEquals(LeadMailbox.DEFAULT_PREFETCH, opener.offeredPrefetch,
+                "an unset prefetch takes the mailbox's own default, not zero");
+    }
+
+    @Test
+    void honoursUriEnvOverALiteralUri() {
+        var opener = new RecordingOpener();
+        var coordinator = new FleetConfig.Coordinator("amqp://stale:stale@old:5672/x", "COORD_URI",
+                "mac-opus", 8);
+
+        Fleetd.openLeadMailbox(coordinator, Map.of("COORD_URI", RESOLVED_URI), opener);
+
+        assertEquals(RESOLVED_URI, opener.offeredUri, "the secret store wins over clear text");
+        assertEquals(8, opener.offeredPrefetch);
+    }
+
+    @Test
+    void turnsOffWhenUriEnvDoesNotResolve() {
+        var opener = new RecordingOpener();
+        var coordinator = new FleetConfig.Coordinator(null, "COORD_URI", "mac-opus", null);
+
+        assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener));
+
+        assertNull(opener.offeredUri);
+    }
+
+    @Test
+    void warnsAndStaysOffWhenSelfIdIsMissing() {
+        var appender = captureFleetdLogs();
+        var opener = new RecordingOpener();
+        var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, null, null);
+
+        assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener));
+
+        assertNull(opener.offeredUri, "a mailbox with no owning coord-id has no queue to declare");
+        String warns = joined(appender, Level.WARN);
+        assertTrue(warns.contains("coordinator.selfId"), () -> "say which key is missing: " + warns);
+        assertFalse(warns.contains(SECRET), () -> "the URI's password must never be logged: " + warns);
+    }
+
+    @Test
+    void warnsAndStaysOffWhenTheBrokerIsUnreachableAtBoot() {
+        var appender = captureFleetdLogs();
+        var opener = new RecordingOpener();
+        opener.unreachable = true;
+        var coordinator = new FleetConfig.Coordinator(RESOLVED_URI, null, "mac-opus", null);
+
+        assertNull(Fleetd.openLeadMailbox(coordinator, Map.of(), opener),
+                "a down coordination broker turns the feature off; it must never take the daemon down");
+
+        String warns = joined(appender, Level.WARN);
+        assertTrue(warns.contains("coord.example"), () -> "name the host that failed: " + warns);
+        assertFalse(warns.contains(SECRET), () -> "with credentials stripped: " + warns);
+        assertTrue(warns.contains("Connection refused"), () -> "and the real reason: " + warns);
+    }
+}

bridged/src/test/java/dev/ltms/fleet/config/FleetConfigTest.java

@@ -43,6 +43,49 @@ class FleetConfigTest {
         assertTrue(cfg.guard().hostSet().contains("ollama.ltms.dev"));
     }
 
+    @Test
+    void aProfileWithAnAutoCompactWindowBelowTheAcceptedRangeIsRejectedAtLoad(@TempDir Path dir)
+            throws Exception {
+        Path f = dir.resolve("low-window.yaml");
+        Files.writeString(f, """
+                profiles:
+                  ltms-local:
+                    baseUrl: http://gx00.gw:8000
+                    autoCompactWindow: 50000
+                """);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class, () -> FleetConfig.load(f));
+        assertTrue(e.getMessage().contains("ltms-local"), "the offending profile is named");
+        assertTrue(e.getMessage().contains("autoCompactWindow"));
+    }
+
+    @Test
+    void aProfileWithAnAutoCompactWindowInRangeLoadsFine(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("in-range-window.yaml");
+        Files.writeString(f, """
+                profiles:
+                  ltms-local:
+                    baseUrl: http://gx00.gw:8000
+                    autoCompactWindow: 250000
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertEquals(250_000, cfg.profiles().get("ltms-local").autoCompactWindow());
+    }
+
+    @Test
+    void aProfileWithNoAutoCompactWindowLeavesItNull(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("no-window.yaml");
+        Files.writeString(f, """
+                profiles:
+                  ltms-local:
+                    baseUrl: http://gx00.gw:8000
+                """);
+
+        assertNull(FleetConfig.load(f).profiles().get("ltms-local").autoCompactWindow(),
+                "unset means off — today's behaviour, unchanged");
+    }
+
     @Test
     void appliesDefaultsForMissingSections(@TempDir Path dir) throws Exception {
         Path f = dir.resolve("minimal.yaml");

bridged/src/test/java/dev/ltms/fleet/mcp/FleetMcpLeadCoordTest.java

@@ -0,0 +1,96 @@
+package dev.ltms.fleet.mcp;
+
+import dev.ltms.fleet.msg.FakeLeadChannel;
+import dev.ltms.fleet.msg.LeadMessage;
+import io.modelcontextprotocol.spec.McpSchema;
+import org.junit.jupiter.api.Test;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+/**
+ * CB-637: the {@code fleet_send{coordId}} route — a message to a PEER LEAD on another daemon,
+ * published to its durable mailbox instead of typed into a pane this daemon can reach.
+ *
+ * <p>Hermetic: {@link FakeLeadChannel} replaces the AMQP-backed {@code LeadMailbox}, so these run
+ * with no broker. What is checked here is routing and refusal — which envelope goes out, and which
+ * calls are rejected before anything is sent.
+ */
+class FleetMcpLeadCoordTest {
+
+    private static final String SELF = "mac-opus";
+    private static final String PEER = "fleet01-lead";
+
+    private static String textOf(McpSchema.CallToolResult r) {
+        return ((McpSchema.TextContent) r.content().getFirst()).text();
+    }
+
+    @Test
+    void publishesAnEnvelopeAddressedFromThisDaemonToThePeer() {
+        var channel = new FakeLeadChannel(SELF);
+
+        McpSchema.CallToolResult res =
+                FleetMcp.sendToLead(channel, PEER, "you own the auth layer, I own the config one", null, null);
+
+        assertFalse(res.isError(), () -> "expected a success result, got: " + textOf(res));
+        assertEquals(1, channel.published().size());
+        LeadMessage sent = channel.published().getFirst();
+        assertEquals(SELF, sent.from(), "the sender is this daemon's own coord-id, never an argument");
+        assertEquals(PEER, sent.to());
+        assertEquals("you own the auth layer, I own the config one", sent.content());
+        assertNotNull(sent.msgId());
+        assertFalse(sent.msgId().isBlank(), "the envelope carries an idempotency id for dedup on redelivery");
+        assertTrue(textOf(res).contains(PEER), "the receipt names the peer it reached");
+    }
+
+    @Test
+    void rejectsCoordIdTogetherWithSessionId() {
+        var channel = new FakeLeadChannel(SELF);
+
+        McpSchema.CallToolResult res = FleetMcp.sendToLead(channel, PEER, "hi", "term_worker", null);
+
+        assertTrue(res.isError());
+        assertTrue(textOf(res).contains("coordId"), () -> textOf(res));
+        assertTrue(textOf(res).contains("sessionId"), () -> "the error must name the conflict: " + textOf(res));
+        assertEquals(0, channel.published().size(), "an ambiguous call must send nothing at all");
+    }
+
+    @Test
+    void rejectsCoordIdTogetherWithTurnId() {
+        var channel = new FakeLeadChannel(SELF);
+
+        McpSchema.CallToolResult res = FleetMcp.sendToLead(channel, PEER, "hi", null, "turn_7");
+
+        assertTrue(res.isError());
+        assertTrue(textOf(res).contains("turnId"), () -> textOf(res));
+        assertEquals(0, channel.published().size());
+    }
+
+    @Test
+    void saysSoPlainlyWhenLeadCoordinationIsNotConfigured() {
+        McpSchema.CallToolResult res = FleetMcp.sendToLead(null, PEER, "hi", null, null);
+
+        assertTrue(res.isError());
+        assertTrue(textOf(res).contains("lead coordination is not configured"), () -> textOf(res));
+        assertTrue(textOf(res).contains("coordinator"), () -> "point at the config block to add: " + textOf(res));
+    }
+
+    @Test
+    void reportsAnUnreachablePeerAsAToolErrorNamingIt() {
+        var channel = new FakeLeadChannel(SELF)
+                .failPublishWith("lead message m1 was returned as unroutable (mailbox not owned)");
+
+        McpSchema.CallToolResult res = FleetMcp.sendToLead(channel, PEER, "hi", null, null);
+
+        assertTrue(res.isError(), "a black-holed message must never be reported as delivered");
+        assertTrue(textOf(res).contains(PEER), () -> "the error must name the coordId: " + textOf(res));
+        assertTrue(textOf(res).contains("unroutable"), () -> "and carry the broker's reason: " + textOf(res));
+    }
+
+    @Test
+    void requiresContent() {
+        var channel = new FakeLeadChannel(SELF);
+
+        assertTrue(FleetMcp.sendToLead(channel, PEER, "   ", null, null).isError());
+        assertEquals(0, channel.published().size());
+    }
+}

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

@@ -527,6 +527,35 @@ class FleetMcpTest {
         assertTrue(out.contains("\"liveStatus\":\"unknown\""), out);
     }
 
+    @Test
+    void listReportsThisDaemonsOwnCoordIdWhenLeadCoordinationIsOn() {
+        FakeHerdr h = new FakeHerdr();
+        SessionManager sessions = new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")));
+
+        McpSchema.CallToolResult res = FleetMcp.listFleet(
+                workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")), sessions, null,
+                FleetMcp.CapacitySource.none(), new FleetMcp.HealthCoverageSource(() -> "off"),
+                FleetMcp.QuarantineSource.none(), Map.of(), "", "mac-opus");
+
+        String out = textOf(res);
+        // There is no peer-discovery surface yet, so this row answers the one question an operator
+        // cannot answer any other way: which coord-id a peer must use to reach ME.
+        assertTrue(out.contains("\"coordinator\""), out);
+        assertTrue(out.contains("\"selfId\":\"mac-opus\""), out);
+    }
+
+    @Test
+    void listOmitsTheCoordinatorRowWhenLeadCoordinationIsOff() {
+        FakeHerdr h = new FakeHerdr();
+        SessionManager sessions = new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")));
+
+        McpSchema.CallToolResult res = FleetMcp.listFleet(
+                workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")), sessions, Map.of(), "");
+
+        assertFalse(textOf(res).contains("coordinator"),
+                "an ordinary fleet's output must be unchanged by this feature");
+    }
+
     @Test
     void capacityUsesThePlacementLiveCount() {
         FakeHerdr h = new FakeHerdr();

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

@@ -1138,6 +1138,42 @@ class ClaudeCodeLauncherTest {
         assertNull(startEnv(herdr).get("ANTHROPIC_MODEL"));
     }
 
+    // ── autoCompactWindow: --autocompact is pinned on the command line, opt-in per profile ───────
+
+    /** A launcher for a profile identical but for its {@code autoCompactWindow:} — the only variable. */
+    private ClaudeCodeLauncher serviceWithAutoCompactWindow(FakeHerdr herdr, Integer window) {
+        FleetConfig.Profile cfg = profileWithAutoCompactWindow(window);
+        return new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(),
+                _ -> null);
+    }
+
+    private static FleetConfig.Profile profileWithAutoCompactWindow(Integer window) {
+        return new FleetConfig.Profile("sonnet", "http://gx00.gw:8000", "claude-sonnet-5", null,
+                "BRIDGED_WORKER_TOKEN", List.of("ccs", "sonnet"), "tab", "bridged-workers",
+                "w #{n}", "http://127.0.0.1:8765/mcp", null, null, null, null, null, Map.of(),
+                null, null, null, null, null, null, null, null, window);
+    }
+
+    @Test
+    void aConfiguredAutoCompactWindowIsPassedAsAnAutocompactFlag() {
+        FakeHerdr herdr = new FakeHerdr();
+        serviceWithAutoCompactWindow(herdr, 250_000).spawn("sonnet", null, null);
+
+        List<String> args = spawnedArgs(herdr);
+        int flag = args.indexOf("--autocompact");
+        assertTrue(flag >= 0, "the flag is what survives a wrapper argv like [ccs, sonnet]");
+        assertEquals("250000", args.get(flag + 1));
+    }
+
+    @Test
+    void aProfileWithNoAutoCompactWindowGetsNoAutocompactFlag() {
+        FakeHerdr herdr = new FakeHerdr();
+        serviceWithAutoCompactWindow(herdr, null).spawn("sonnet", null, null);
+
+        assertFalse(spawnedArgs(herdr).contains("--autocompact"));
+    }
+
     // --- CB-539: subscription-profile opt-in ----------------------------------------------------
 
     /** A claude-code profile on the subscription: no baseUrl (by design), no off-sub endpoint. */

bridged/src/test/java/dev/ltms/fleet/member/OpenCodeLauncherTest.java

@@ -495,6 +495,71 @@ class OpenCodeLauncherTest {
                 "without a baseUrl opencode resolves its own provider as before");
     }
 
+    // --- autoCompactWindow: opencode has no absolute compact-at-N knob, so this is applied as the
+    //     model's own limit.context, only when model: resolves to "provider/model" -----------------
+
+    private static FleetConfig.Profile opencodeCfgWithAutoCompactWindow(String model, String baseUrl,
+                                                                        Integer window) {
+        return new FleetConfig.Profile("gemini", baseUrl, model, null, "BRIDGED_WORKER_TOKEN",
+                List.of("opencode"), "tab", "bridged-workers", "opencode: {model} #{n}", null,
+                null, null, null, null, FleetConfig.Profile.KIND_OPENCODE, Map.of(), null, null,
+                null, null, null, null, null, null, window);
+    }
+
+    @Test
+    void autoCompactWindowIsAppliedAsThePerModelContextLimit(@TempDir Path root) throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        service(herdr, root, opencodeCfgWithAutoCompactWindow("openai/gpt-5", null, 250_000)).spawn();
+
+        String cfgPath = startEnv(herdr).get("OPENCODE_CONFIG");
+        assertNotNull(cfgPath, "autoCompactWindow alone must trigger config generation, with no MCP"
+                + " and no custom provider set");
+        JsonNode limit = new ObjectMapper().readTree(Path.of(cfgPath).toFile())
+                .path("provider").path("openai").path("models").path("gpt-5").path("limit");
+        assertEquals(250_000, limit.path("context").asInt());
+        assertEquals(16384, limit.path("output").asInt(),
+                "opencode's limit schema requires both keys; output gets a safe documented default");
+    }
+
+    @Test
+    void autoCompactWindowMergesIntoACustomProviderRatherThanOverwritingIt(@TempDir Path root)
+            throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        service(herdr, root, opencodeCfgWithAutoCompactWindow(
+                "local-vllm/deepseek-v4-flash", "http://127.0.0.1:8000", 300_000)).spawn();
+
+        JsonNode provider = new ObjectMapper()
+                .readTree(Path.of(startEnv(herdr).get("OPENCODE_CONFIG")).toFile())
+                .path("provider").path("local-vllm");
+        assertEquals("@ai-sdk/openai-compatible", provider.path("npm").asText(),
+                "addCustomProvider's own fields must survive the later limit merge");
+        assertEquals(300_000, provider.path("models").path("deepseek-v4-flash")
+                .path("limit").path("context").asInt());
+        assertEquals("deepseek-v4-flash", provider.path("models").path("deepseek-v4-flash")
+                .path("name").asText(),
+                "the model's pre-existing 'name' field must survive the limit merge too");
+    }
+
+    @Test
+    void autoCompactWindowWithNoProviderSlashInModelGetsNoLimitAndAWarn(@TempDir Path root)
+            throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        // mcpUrl set too, only so a config file gets written at all to inspect; a bare model name
+        // with no other config-triggering knob would leave OPENCODE_CONFIG unset entirely, which is
+        // also correct (nothing to write) but not what this test is asserting.
+        service(herdr, root, new FleetConfig.Profile("gemini", null, "some-free-model", null,
+                "BRIDGED_WORKER_TOKEN", List.of("opencode"), "tab", "bridged-workers",
+                "opencode: {model} #{n}", "http://127.0.0.1:8765/mcp", null, null, null, null,
+                FleetConfig.Profile.KIND_OPENCODE, Map.of(), null, null, null, null, null, null,
+                null, null, 250_000)).spawn();
+
+        String cfgPath = startEnv(herdr).get("OPENCODE_CONFIG");
+        JsonNode json = new ObjectMapper().readTree(Path.of(cfgPath).toFile());
+        assertTrue(json.path("provider").isMissingNode(),
+                "a bare model name cannot be targeted at a specific provider/model limit entry — "
+                        + "no silent no-op, but also no broken partial write");
+    }
+
     // --- CB-634: IDE Index MCP + guidance overlay (opencode does not read CLAUDE.local.md) -------
 
     private static FleetConfig.Profile opencodeIdeCfg(String mcpUrl, String ideUrl, String cwd) {

bridged/src/test/java/dev/ltms/fleet/msg/FakeLeadChannel.java

@@ -0,0 +1,75 @@
+package dev.ltms.fleet.msg;
+
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.List;
+
+/**
+ * Hermetic stand-in for {@link LeadChannel}: an in-memory mailbox that records what was published
+ * and what was acked, with no broker anywhere.
+ *
+ * <p>Its whole point is that {@link LeadMailbox} — the production implementation — owns a live AMQP
+ * connection, so every test of the code AROUND it would otherwise need a broker and end up tagged
+ * {@code contract}. The broker round trip is covered once, by {@code LeadMailboxTest}; the routing
+ * ({@code FleetMcp}) and the delivery ({@link LeadCoordLoop}) are covered here.
+ *
+ * <p>Thread-safe: {@link LeadCoordLoop} calls it from its own scheduler thread while a test reads
+ * the recorded lists.
+ */
+public final class FakeLeadChannel implements LeadChannel {
+
+    private final String selfCoordId;
+    private final List<LeadMessage> held = Collections.synchronizedList(new ArrayList<>());
+    private final List<LeadMessage> published = Collections.synchronizedList(new ArrayList<>());
+    private final List<String> acked = Collections.synchronizedList(new ArrayList<>());
+    /** When set, every {@link #publish} throws it — the unroutable/nacked/timed-out peer. */
+    private volatile IllegalStateException publishFailure;
+
+    public FakeLeadChannel(String selfCoordId) {
+        this.selfCoordId = selfCoordId;
+    }
+
+    /** Make every publish fail as an unreachable peer would. */
+    public FakeLeadChannel failPublishWith(String message) {
+        this.publishFailure = new IllegalStateException(message);
+        return this;
+    }
+
+    /** Put a message in this mailbox as if a peer had sent it. */
+    public FakeLeadChannel hold(LeadMessage m) {
+        held.add(m);
+        return this;
+    }
+
+    @Override
+    public void publish(String toCoordId, LeadMessage m) {
+        if (publishFailure != null) {
+            throw publishFailure;
+        }
+        published.add(m);
+    }
+
+    @Override
+    public List<LeadMessage> peek() {
+        return List.copyOf(held);
+    }
+
+    @Override
+    public void ack(String msgId) {
+        held.removeIf(m -> m.msgId().equals(msgId));
+        acked.add(msgId);
+    }
+
+    @Override
+    public String selfCoordId() {
+        return selfCoordId;
+    }
+
+    public List<LeadMessage> published() {
+        return List.copyOf(published);
+    }
+
+    public List<String> acked() {
+        return List.copyOf(acked);
+    }
+}

bridged/src/test/java/dev/ltms/fleet/msg/LeadCoordLoopTest.java

@@ -0,0 +1,150 @@
+package dev.ltms.fleet.msg;
+
+import dev.ltms.fleet.herdr.AgentControl;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import org.junit.jupiter.api.Test;
+
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.ScheduledExecutorService;
+
+import static org.junit.jupiter.api.Assertions.*;
+
+/**
+ * Unit tests for {@link LeadCoordLoop} — the receive half of lead-to-lead messaging. Hermetic
+ * throughout: a {@link FakeLeadChannel} stands in for the mailbox and {@link FakeHerdr} for the
+ * pane, so no broker and no herdr daemon is involved. The broker round trip has its own
+ * {@code contract}-tagged test on {@link LeadMailbox}.
+ *
+ * <p>Every test drives {@link LeadCoordLoop#tick()} directly rather than waiting on the scheduler:
+ * the schedule itself is one {@code scheduler.schedule} call, while the decisions worth pinning —
+ * inject or hold, ack or leave unacked — all live in the tick.
+ */
+class LeadCoordLoopTest {
+
+    private static final String SELF = "mac-opus";
+    private static final String PEER = "fleet01-lead";
+    private static final String LEAD_TERM = "term_lead";
+
+    /** No scheduler is needed: nothing here calls start(). */
+    private static final ScheduledExecutorService NO_SCHEDULER = null;
+
+    private static LeadCoordLoop loop(LeadChannel channel, FakeHerdr herdr, Map<String, String> leads) {
+        return new LeadCoordLoop(channel, new AgentControl(herdr), () -> leads, NO_SCHEDULER, 3_000L);
+    }
+
+    private static List<FakeHerdr.Call> prompts(FakeHerdr herdr) {
+        return herdr.calls.stream().filter(c -> c.method().equals("agent.prompt")).toList();
+    }
+
+    @Test
+    void deliversAHeldMessageToTheLeadPaneAndAcksIt() {
+        var channel = new FakeLeadChannel(SELF).hold(new LeadMessage("m1", PEER, SELF, "the merge is blocked"));
+        var herdr = new FakeHerdr().agentStatus("idle");
+
+        loop(channel, herdr, Map.of(LEAD_TERM, SELF)).tick();
+
+        assertEquals(1, prompts(herdr).size(), "an injectable lead must receive the peer's message");
+        @SuppressWarnings("unchecked")
+        Map<String, Object> params = (Map<String, Object>) prompts(herdr).getFirst().params();
+        assertEquals(LEAD_TERM, params.get("target"), "delivered to the resolved local lead pane");
+        assertEquals("[lead " + PEER + "] the merge is blocked", params.get("text"),
+                "the sender's coord-id is carried into the pane — the lead must know who to answer");
+        assertEquals(List.of("m1"), channel.acked(), "a delivered message is acked off the broker");
+        assertTrue(channel.peek().isEmpty(), "and is no longer held");
+    }
+
+    @Test
+    void leavesTheMessageUnackedWhenTheLeadIsMidTurn() {
+        var channel = new FakeLeadChannel(SELF).hold(new LeadMessage("m1", PEER, SELF, "hello"));
+        var herdr = new FakeHerdr().agentStatus("working");
+
+        loop(channel, herdr, Map.of(LEAD_TERM, SELF)).tick();
+
+        assertEquals(0, prompts(herdr).size(), "never paste into a live turn");
+        assertEquals(List.of(), channel.acked(), "an undelivered message must NOT be acked");
+        assertEquals(1, channel.peek().size(), "it stays held for the next tick");
+    }
+
+    @Test
+    void leavesTheMessageUnackedWhenNoLeadPaneIsKnown() {
+        var channel = new FakeLeadChannel(SELF).hold(new LeadMessage("m1", PEER, SELF, "hello"));
+        var herdr = new FakeHerdr().agentStatus("idle");
+
+        loop(channel, herdr, Map.of()).tick();
+
+        assertEquals(0, prompts(herdr).size());
+        assertEquals(List.of(), channel.acked(), "nowhere to deliver is not a reason to drop it");
+        assertEquals(1, channel.peek().size());
+    }
+
+    @Test
+    void leavesTheMessageUnackedWhenHerdrRefusesTheInjection() {
+        var channel = new FakeLeadChannel(SELF).hold(new LeadMessage("m1", PEER, SELF, "hello"));
+        var herdr = new FakeHerdr().agentStatus("idle").agentSendFailsWith("agent_not_found");
+
+        loop(channel, herdr, Map.of(LEAD_TERM, SELF)).tick();
+
+        assertEquals(List.of(), channel.acked(), "a send that threw delivered nothing, so nothing is acked");
+        assertEquals(1, channel.peek().size());
+    }
+
+    @Test
+    void resolvesTheLeadByNameWhenSeveralAreKnown() {
+        var channel = new FakeLeadChannel(SELF).hold(new LeadMessage("m1", PEER, SELF, "hello"));
+        var herdr = new FakeHerdr().agentStatus("idle");
+        // Two leads on this daemon; only one carries the coord-id the mailbox is owned as.
+        var leads = new java.util.LinkedHashMap<String, String>();
+        leads.put("term_other", "some-other-lead");
+        leads.put(LEAD_TERM, SELF);
+
+        loop(channel, herdr, leads).tick();
+
+        @SuppressWarnings("unchecked")
+        Map<String, Object> params = (Map<String, Object>) prompts(herdr).getFirst().params();
+        assertEquals(LEAD_TERM, params.get("target"), "the lead named after coordinator.selfId wins");
+    }
+
+    @Test
+    void holdsWhenSeveralLeadsAreKnownAndNoneCarriesTheCoordId() {
+        var channel = new FakeLeadChannel(SELF).hold(new LeadMessage("m1", PEER, SELF, "hello"));
+        var herdr = new FakeHerdr().agentStatus("idle");
+        var leads = new java.util.LinkedHashMap<String, String>();
+        leads.put("term_one", "lead-one");
+        leads.put("term_two", "lead-two");
+
+        loop(channel, herdr, leads).tick();
+
+        assertEquals(0, prompts(herdr).size(),
+                "guessing a pane would type a peer's message into the wrong lead and then ack it");
+        assertEquals(List.of(), channel.acked());
+    }
+
+    @Test
+    void deliversOneMessagePerTickSoEachIsGatedOnItsOwnStatusRead() {
+        var channel = new FakeLeadChannel(SELF)
+                .hold(new LeadMessage("m1", PEER, SELF, "first"))
+                .hold(new LeadMessage("m2", PEER, SELF, "second"));
+        var herdr = new FakeHerdr().agentStatus("idle");
+        var loop = loop(channel, herdr, Map.of(LEAD_TERM, SELF));
+
+        loop.tick();
+        assertEquals(List.of("m1"), channel.acked(), "FIFO: the oldest goes first");
+        assertEquals(1, prompts(herdr).size(), "the second waits for a fresh status read");
+
+        loop.tick();
+        assertEquals(List.of("m1", "m2"), channel.acked());
+        assertEquals(2, prompts(herdr).size());
+        assertTrue(channel.peek().isEmpty());
+    }
+
+    @Test
+    void anEmptyMailboxNeverTouchesHerdr() {
+        var channel = new FakeLeadChannel(SELF);
+        var herdr = new FakeHerdr().agentStatus("idle");
+
+        loop(channel, herdr, Map.of(LEAD_TERM, SELF)).tick();
+
+        assertEquals(0, herdr.calls.size(), "an idle fleet must not poll a pane's status every tick");
+    }
+}