CB-604: reject an unrecognized profile kind at config load

Most of issue #65 was already shipped in 5d5b3bd - MemberSession
records agentSessionId, PeerHandle.agentSessionId() has no default, the
roster exposes it, and bridge_spawn accepts sessionName and
resumeSessionId. That commit left one item for follow-up: carrying the
id on ReleaseDetail.

This does that item. CB-578 stage C already lets a lead re-dispatch onto
the same worktree after a failure; without the session id that is a cold
start. With it, the work and the thread both survive.

Also updates the bridge_spawn and bridge_list rows in the CLAUDE.md
intent table, which the earlier commit missed.

Verified here: trial merge onto main builds 830 tests BUILD SUCCESS,
unpiped. Confirmed against the code that items 1-4 really were already
on main, so the scope-down is correct rather than work skipped.

Closes #65

CLAUDE.md

@@ -105,8 +105,8 @@ the merge — and merging on a reviewer's word is delegating it by proxy.
 |---|---|
 | Confirm your own role | `bridge_whoami` |
 | See backends available | `bridge_profiles` |
-| Start a member | `bridge_spawn{role?, profile?, cwd?, worktree?, ticket?}` → `sessionId` + `paneId` |
-| See the fleet | `bridge_list` → `leads` (your peers) + `members` · one peer's state: `bridge_status{sessionId}` |
+| Start a member | `bridge_spawn{role?, profile?, cwd?, worktree?, ticket?, sessionName?, resumeSessionId?}` → `sessionId` + `paneId` |
+| See the fleet | `bridge_list` → `leads` (your peers) + `members` (each carries `agentSessionId` when its backend knows one) · one peer's state: `bridge_status{sessionId}` |
 | Delegate (blocking) | `bridge_send{sessionId, content}` |
 | Delegate (long task) | `bridge_send{sessionId, content, wait:false}` → ticket → `bridge_poll{ticket}` |
 | Answer a member's `bridge_ask` | `bridge_send{turnId, content}` — **not** `sessionId` |

bridged/bridged.example.yaml

@@ -88,15 +88,28 @@ bind:
 #   backoffMs: 60000
 #   quietNudgeCap: 3
 
-# Fleet health detection is dormant unless enabled. It reads one whole-fleet agent list per tick.
-# It can run without a webhook; bridge_list then reports healthCoverage: detection-only.
+# Fleet health detection is dormant unless enabled (CB-573). It reads one whole-fleet agent list
+# per tick.
+#   intervalSeconds → how often a tick runs (default 30). ENFORCED floor of 15: the code computes
+#                      Math.max(15, intervalSeconds), so a lower value is silently raised, not
+#                      rejected.
+#   workingSuspectAfterSeconds, paneProbeIntervalSeconds → accepted and parsed, but NOT YET READ by
+#                      anything — the dormant monitor only consumes intervalSeconds today (CB-573
+#                      shipped ahead of the evidence publishers these two knobs are for). Setting
+#                      them changes nothing right now, and no minimum is enforced on either, because
+#                      nothing reads them to enforce one. They exist so a later build can start
+#                      honouring them without another config-shape change.
+#   notifications.mode → "webhook" flips what bridge_list REPORTS (healthCoverage: "full" instead
+#                      of "detection-only") — it does NOT make bridged send any webhook call; no
+#                      delivery mechanism is implemented yet. Any other value, or omitting the
+#                      block, reports "detection-only".
 # health:
 #   enabled: true
-#   intervalSeconds: 30              # minimum 15
-#   workingSuspectAfterSeconds: 600  # minimum 300
-#   paneProbeIntervalSeconds: 60     # minimum 60
+#   intervalSeconds: 30
+#   workingSuspectAfterSeconds: 600
+#   paneProbeIntervalSeconds: 60
 #   notifications:
-#     mode: disabled                 # disabled (default) or webhook
+#     mode: disabled
 
 # herdr Unix socket. Omit to use the client default
 # (${HERDR_SOCKET_PATH:-~/.config/herdr/herdr.sock}).
@@ -262,6 +275,26 @@ profiles:
   #     argv: ["opencode"]
 # How an unqualified spawn chooses a profile: fixed (default, reproduces pre-CB-518 behaviour),
 # round-robin, or weighted. Omitting this key is a strict no-op for existing configs.
+#
+# `weighted` IS NOT "cheapest first" — read this before you set weights (CB-589).
+# It is smooth weighted round-robin: it spreads spawns across EVERY profile that has a free slot,
+# in weight ratio. It has no idea which profile costs money. So with local:10 / paid:2 you do not
+# get "use local, overflow to paid" — you get roughly one spawn in six going to the paid profile
+# while the local box still has a free slot.
+#
+# There is a sharper second effect. The policy's running score map lives for the daemon's whole
+# life. While a profile is at maxLoad it is filtered out and its score FREEZES, so the paid
+# profiles keep accumulating against it. When the local slot frees up it returns with a stale
+# score and can LOSE the next pick — a paid spawn while the free box sits idle.
+#
+# Until a real cost-first policy exists, the workaround is to make the ratio decisive rather than
+# proportional: give the free profile a weight so large that it wins every pick it is eligible
+# for, and paid profiles only ever take genuine overflow. On this host that is local weight 100
+# against paid weights of ~1.
+#
+# The gotcha with that workaround: it expresses a PREFERENCE ORDER through a RATIO knob. Add a
+# future profile at weight 150 and it silently outranks the free box, with nothing to warn you.
+# Re-check the weights whenever you add a profile.
 placement: weighted
 
 # How long a credential sits out after a BACKEND_EXHAUSTED classification (CB-578 stage B), in
@@ -286,6 +319,11 @@ placement: weighted
 #               / credentialId. Those are hot because the placement policy (and, for credentialId,
 #               the CB-578 stage B quarantine check) reads them through a supplier — being config is
 #               not by itself enough to make a key hot.
+#               EXCEPT `fleet.leaders`: Bridged.main reads it once at startup to build the lead tab
+#               scanner and launcher, and neither is rebuilt on reload. A changed/added/removed
+#               `fleet.leaders` entry is silently accepted — the reload reports "config reloaded"
+#               with nothing in the deferred list — but has NO effect until you restart. Treat it
+#               as deferred in practice, even though today's reload output does not say so.
 #   DEFERRED  → accepted into the new config, but the wiring built at startup keeps the old value
 #               until you restart: `lifecycle:`, `leadHeartbeat:`, `guard:`, `worktreeRoot:`,
 #               `spawnReadyTimeoutMs` / `spawnReadyPollMs`, `quarantineCooldownSeconds` (CB-578
@@ -368,6 +406,12 @@ fleet:
   # An auto-launched lead is NOT a member: it gets no worker reply charter, is never registered with
   # the session lifecycle (the idle reaper would kill your orchestrator), and stays on the
   # subscription — ANTHROPIC_BASE_URL/AUTH_TOKEN are stripped from its env whatever the profile says.
+  #
+  # GET THE `tab:` VALUE RIGHT. A pane that does not match any configured `tab:` (a typo, a renamed
+  # tab, a pane no entry names at all) is not recognised as a lead — it resolves as an ordinary
+  # WORKER instead, silently, and every orchestration call it makes (spawn/stop/send/drain) is
+  # refused. There is no error at startup for this: an unmatched pane is simply not a lead. If your
+  # primary suddenly can't spawn or send, check this section first.
   # leaders:
   #   opus-5.0:
   #     profile: opus                # omit to never create this lead, only recognise it
@@ -423,10 +467,15 @@ guard:
 #   idleTtlSeconds     → reap READY/DONE sessions idle longer than this (never BUSY/SPAWNING)
 #   contextCap         → force-release a session after this many delegated turns
 #   drainTimeoutSeconds → seconds to wait for BUSY sessions on shutdown before forced teardown
+#   clearAfterTurn     → whether a reusable worker discards its conversation context after every
+#                        completed delegated turn (default false). Works for claude-code workers
+#                        only — any other peer kind (e.g. opencode) logs "context reset is
+#                        unsupported for peer kind …" once and the reset is a no-op.
 # lifecycle:
 #   idleTtlSeconds: 300
 #   contextCap: 10
 #   drainTimeoutSeconds: 5
+#   clearAfterTurn: false
 
 # Durable reply delivery (CB-307 Stage 2). OMIT this block entirely to keep the default
 # in-memory, soft-state reply inbox (late worker replies are held only until a daemon bounce).

bridged/src/main/java/dev/ltms/bridged/Bridged.java

@@ -441,6 +441,11 @@ public final class Bridged {
                 reason += "; worktree=" + detail.worktreePath() + " branch=" + detail.branch()
                         + " snapshot=" + (detail.snapshotRef() != null ? detail.snapshotRef() : "none");
             }
+            // CB-584 (issue #65 criterion 5): also name the agent session, so a lead can resume the
+            // member's conversation instead of only re-dispatching a fresh one onto the same files.
+            if (detail.agentSessionId() != null) {
+                reason += " agentSessionId=" + detail.agentSessionId();
+            }
             messages.abandon(detail.terminalId(), reason);
             replyInbox.release(detail.terminalId());
             primaryRegistry.forgetDelegation(detail.terminalId()); // CB-532: don't leak the lead binding
@@ -555,7 +560,10 @@ public final class Bridged {
      * where set (opt-in). Derived from the config, not hard-coded, so a new profile is covered for
      * free. A var required by more than one profile is one entry naming every profile that needs
      * it. Deliberately excludes {@code auth.tokenEnv}: that one is already enforced loudly, by a
-     * startup throw, a few lines above this method's call site.
+     * startup throw in {@code main()} — about 370 lines <em>below</em> this method's call site
+     * ({@link #reportRequiredSecrets(BridgedConfig)}), not a few lines above it. That throw only
+     * fires when {@code auth.mode: token} is configured; under the default loopback-trust mode it
+     * never runs, and {@code auth.tokenEnv} is simply not required.
      *
      * <p>Package-private and pure (no I/O, no logging) so the derivation is unit-testable without
      * capturing log output; {@link #reportRequiredSecrets(BridgedConfig)} is the logging caller.

bridged/src/main/java/dev/ltms/bridged/config/BridgedConfig.java

@@ -967,8 +967,12 @@ public record BridgedConfig(
     /**
      * Top-level keys this version understands. Used only to warn about the rest — see
      * {@link #warnUnknownTopLevelKeys}. Keep in step with the record components.
+     *
+     * <p>Package-private (not {@code private}) so a test can assert every key here is documented in
+     * {@code bridged.example.yaml} — the only committed description of the config schema, since
+     * {@code bridged.yaml} itself is gitignored.
      */
-    private static final Set<String> KNOWN_TOP_LEVEL_KEYS = Set.of(
+    static final Set<String> KNOWN_TOP_LEVEL_KEYS = Set.of(
             "bind", "herdrSocket", "profiles", "guard", "worktreeRoot",
             "lifecycle", "spawnReadyTimeoutMs", "spawnReadyPollMs", "broker", "primary", "fleet",
             "leadHeartbeat", "health", "placement", "auth", "configReload", "quarantineCooldownSeconds");
@@ -982,6 +986,7 @@ public record BridgedConfig(
             warnUnknownTopLevelKeys(yaml, path);
             rejectDuplicateMemberSlots(yaml);
             rejectNegativeMaxLoad(yaml);
+            rejectUnknownKind(yaml);
             BridgedConfig cfg = YAML.readValue(yaml, BridgedConfig.class);
             return cfg.withDefaults();
         } catch (IOException e) {
@@ -1276,6 +1281,53 @@ public record BridgedConfig(
         }
     }
 
+    /** 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);
+
+    /**
+     * Reject a profile whose {@code kind:} is not one of {@link #KNOWN_KINDS} (CB-604), naming the
+     * profile, the value it set, and the accepted set.
+     *
+     * <p>{@link Profile}'s compact constructor only lower-cases {@code kind} and compares it against
+     * {@code KIND_OPENCODE} — anything else, including a typo like {@code opencod}, silently falls
+     * into the claude-code bucket ({@link dev.ltms.bridged.member.CompositePeerLauncher} routes by
+     * exact adapter claim, not by membership in a known set). With {@code argv:} also unset, the argv
+     * default special-cases only the exact string {@code "claude-code"}, so the launch command falls
+     * back to {@code List.of(kind)} — the daemon then tries to run a program literally named after the
+     * typo. {@code CompositePeerLauncher}'s constructor already treats a profile claimed by two
+     * adapters as fatal (CB-402); an unrecognized kind is the same class of adapter-routing mistake
+     * and gets the same treatment here, at config load, rather than surfacing later as a failed spawn.
+     *
+     * @param yaml the raw config text
+     * @throws IllegalStateException when any profile's {@code kind} is a non-blank value not in
+     *                                {@link #KNOWN_KINDS} (case-insensitive)
+     */
+    static void rejectUnknownKind(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("kind") instanceof String k && !k.isBlank()
+                        && !KNOWN_KINDS.contains(k.toLowerCase()))
+                .map(e -> String.valueOf(e.getKey()) + "=" + ((Map<?, ?>) e.getValue()).get("kind"))
+                .sorted()
+                .toList();
+        if (!bad.isEmpty()) {
+            throw new IllegalStateException("refusing to start: profile(s) [" + String.join(", ", bad)
+                    + "] set an unrecognized kind — accepted values are "
+                    + String.join(", ", KNOWN_KINDS.stream().sorted().toList())
+                    + " (case-insensitive); an unrecognized kind would otherwise fall back to the"
+                    + " claude-code adapter and try to launch a program named after the typo.");
+        }
+    }
+
     static List<String> unknownTopLevelKeys(String yaml) {
         Map<?, ?> raw;
         try {

bridged/src/main/java/dev/ltms/bridged/mcp/BridgeMcp.java

@@ -15,6 +15,7 @@ import dev.ltms.bridged.msg.MessageService;
 import dev.ltms.bridged.msg.Rendezvous;
 import dev.ltms.bridged.peer.PeerUnreachableException;
 import dev.ltms.bridged.placement.BackendQuarantine;
+import dev.ltms.bridged.placement.PlacementException;
 import dev.ltms.bridged.session.SessionManager;
 import dev.ltms.bridged.session.MemberSession;
 import dev.ltms.bridged.session.WorktreeRequest;
@@ -692,6 +693,10 @@ public final class BridgeMcp {
             return text(json(memberView(member)));
         } catch (GuardException e) {
             return error("subscription boundary: " + e.getMessage());
+        } catch (PlacementException e) {
+            // CB-599: no candidate had capacity (maxLoad, quarantine, or all-exhausted) — distinct
+            // from "profile does not exist" below.
+            return error("no capacity: " + e.getMessage());
         } catch (IllegalArgumentException e) {
             return error(e.getMessage()); // unknown / no-default profile, or a refused resumeSessionId
         } catch (PeerUnreachableException e) {

bridged/src/main/java/dev/ltms/bridged/rest/BridgedApp.java

@@ -13,6 +13,7 @@ import dev.ltms.bridged.herdr.HerdrClient;
 import dev.ltms.bridged.herdr.HerdrException;
 import dev.ltms.bridged.inject.MemberPresence;
 import dev.ltms.bridged.peer.PeerUnreachableException;
+import dev.ltms.bridged.placement.PlacementException;
 import dev.ltms.bridged.msg.MessageService;
 import dev.ltms.bridged.session.SessionManager;
 import dev.ltms.bridged.peer.MemberRole;
@@ -292,6 +293,11 @@ public final class BridgedApp {
             ctx.status(201).json(view(member));
         } catch (GuardException e) {
             ctx.status(403).json(Map.of("error", "subscription_boundary", "detail", e.getMessage()));
+        } catch (PlacementException e) {
+            // CB-599: no candidate had capacity (maxLoad, quarantine, or all-exhausted) — a benign,
+            // likely-transient refusal, distinct from "profile does not exist" below. 503: the
+            // request was valid and will likely succeed later.
+            ctx.status(503).json(Map.of("error", "no_capacity", "detail", e.getMessage()));
         } catch (IllegalArgumentException e) {
             ctx.status(400).json(Map.of("error", "unknown_profile", "detail", e.getMessage()));
         } catch (PeerUnreachableException e) {

bridged/src/main/java/dev/ltms/bridged/session/SessionManager.java

@@ -293,8 +293,10 @@ public final class SessionManager implements TurnListener {
                 // blocked caller fails fast with a real reason instead of sitting on a rendezvous
                 // nothing will ever resolve. CB-578 stage C: carry the worktree/branch/snapshot ref
                 // too, so a failed ticket's detail can point a lead at the same tree to re-dispatch.
+                // CB-584 (issue #65 criterion 5): carry agentSessionId alongside them, so a lead can
+                // also resume the member's conversation, not just re-dispatch onto its files.
                 notifyReleased(new ReleaseDetail(removed.terminalId(), removed.worktree(),
-                        removed.branch(), snapshotRef));
+                        removed.branch(), snapshotRef, removed.agentSessionId()));
             }
         }
         // CB-581: the pane must always stop, even if the dirty check above threw. A session removed
@@ -348,7 +350,8 @@ public final class SessionManager implements TurnListener {
      * are {@code null} for a shared-tree session; {@code snapshotRef} is {@code null} unless this
      * release snapshotted a dirty worktree into {@code refs/wip/<branch>}.
      */
-    public record ReleaseDetail(String terminalId, String worktreePath, String branch, String snapshotRef) {
+    public record ReleaseDetail(String terminalId, String worktreePath, String branch, String snapshotRef,
+                                String agentSessionId) {
     }
 
     /**

bridged/src/test/java/dev/ltms/bridged/config/BridgedConfigTest.java

@@ -10,6 +10,7 @@ import java.nio.file.Path;
 import java.util.List;
 import java.util.Map;
 import java.util.Set;
+import java.util.regex.Pattern;
 
 import static org.junit.jupiter.api.Assertions.*;
 
@@ -1039,6 +1040,44 @@ class BridgedConfigTest {
                 "an opencode worker with no argv defaults to the opencode binary, never claude");
     }
 
+    /**
+     * CB-604: an unrecognized {@code kind:} used to silently fall into the claude-code bucket — not
+     * matching {@code "opencode"} was the only check. With {@code argv:} also unset that meant the
+     * daemon tried to launch a program literally named after the typo.
+     */
+    @Test
+    void unknownKindIsRefusedAtLoadNamingTheValueAndTheAcceptedSet(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("kind-typo.yaml");
+        Files.writeString(f, """
+                profiles:
+                  gemini:
+                    kind: opencod
+                    model: google/gemini-2.5-pro
+                """);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class, () -> BridgedConfig.load(f));
+        assertTrue(e.getMessage().contains("gemini"), "error names the profile: " + e.getMessage());
+        assertTrue(e.getMessage().contains("opencod"), "error names the bad value: " + e.getMessage());
+        assertTrue(e.getMessage().contains("claude-code") && e.getMessage().contains("opencode"),
+                "error names the accepted set: " + e.getMessage());
+    }
+
+    @Test
+    void blankKindStillDefaultsToClaudeCode(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("kind-blank.yaml");
+        Files.writeString(f, """
+                profiles:
+                  gx10:
+                    baseUrl: http://gx10.gw:8000
+                    kind: ""
+                    argv: ["ccs", "gx10"]
+                """);
+
+        BridgedConfig cfg = BridgedConfig.load(f);
+        assertEquals(BridgedConfig.Profile.KIND_CLAUDE_CODE, cfg.profiles().get("gx10").kind(),
+                "a blank kind: is documented to behave exactly like an absent one");
+    }
+
     @Test
     void authDefaultsToLoopbackTrustSoExistingConfigsBehaveAsBefore(@TempDir Path dir) throws Exception {
         Path f = dir.resolve("no-auth-block.yaml");
@@ -1189,6 +1228,79 @@ class BridgedConfigTest {
         assertEquals(5, cfg.leadHeartbeat().quietNudgeCap());
     }
 
+    /**
+     * A top-level key {@code BridgedConfig} reads but that appears nowhere in
+     * {@code bridged.example.yaml} — live or commented — is invisible drift: {@code bridged.yaml}
+     * is gitignored, so the example is the ONLY committed description of the config schema, and
+     * neither {@link #shippedExampleConfigParses} (example → code: does the example still parse)
+     * nor {@link #everyOptionalKnobDocumentedInTheExampleBinds} (a hand-maintained list of keys
+     * that must bind) can catch a brand-new key nobody added to either.
+     *
+     * <p>This test compares the OTHER direction: every key in {@link BridgedConfig#KNOWN_TOP_LEVEL_KEYS}
+     * (the parser's own accepted set, which backs the unknown-key WARN) must appear as a top-level
+     * key in the example text, live or commented-out — see {@link #topLevelKeyDocumented}.
+     */
+    @Test
+    void everyKnownTopLevelKeyIsDocumentedInTheExample() throws Exception {
+        Path example = Path.of("bridged.example.yaml");
+        assertTrue(Files.exists(example), "bridged.example.yaml must ship next to the pom");
+        String text = Files.readString(example);
+
+        List<String> undocumented = BridgedConfig.KNOWN_TOP_LEVEL_KEYS.stream()
+                .filter(key -> !topLevelKeyDocumented(text, key))
+                .sorted()
+                .toList();
+
+        assertTrue(undocumented.isEmpty(), () -> "key(s) " + undocumented
+                + " are read by BridgedConfig but appear nowhere in bridged.example.yaml — "
+                + "document each one there, commented out if optional. bridged.yaml is "
+                + "gitignored, so this file is the only committed description of the config "
+                + "schema an operator or a worker can see.");
+    }
+
+    /**
+     * Most of {@code bridged.example.yaml} is deliberately commented out — optional sections are
+     * documented as commented blocks so the shipped file stays a working minimal config. A key
+     * documented ONLY as a comment must still count as documented; parsing the file as YAML and
+     * reading its live key set (as an earlier attempt at this guard did) gets this wrong, because
+     * every commented section then looks entirely absent.
+     */
+    @Test
+    void commentedOnlyTopLevelKeyCountsAsDocumented() {
+        String yaml = """
+                bind:
+                  port: 8765
+                # broker:
+                #   uri: amqp://guest:guest@127.0.0.1:5672
+                """;
+        assertTrue(topLevelKeyDocumented(yaml, "broker"),
+                "a key documented only inside a commented-out block must still count as documented");
+    }
+
+    /** A key that appears in neither a live nor a commented top-level line must NOT count. */
+    @Test
+    void absentTopLevelKeyIsNotDocumented() {
+        String yaml = """
+                bind:
+                  port: 8765
+                """;
+        assertFalse(topLevelKeyDocumented(yaml, "broker"),
+                "a key mentioned nowhere in the example must not be reported as documented");
+    }
+
+    /**
+     * True when {@code key} appears as a top-level YAML key in {@code yaml} — either live
+     * ({@code key:} at column 0) or commented out ({@code # key:}, also at column 0, with only
+     * whitespace between the {@code #} and the key). Anchoring on column 0 is what keeps this a
+     * top-level check: an indented occurrence (a nested field, or prose inside a comment that
+     * happens to end in a colon) never matches, because {@code ^} requires the key's own first
+     * character — or the sole leading {@code #} — to sit at the very start of the line.
+     */
+    private static boolean topLevelKeyDocumented(String yaml, String key) {
+        Pattern p = Pattern.compile("(?m)^(?:#\\s*)?" + Pattern.quote(key) + ":");
+        return p.matcher(yaml).find();
+    }
+
     @Test
     void placementDefaultsToFixedForExistingConfigs(@TempDir Path dir) throws Exception {
         Path f = dir.resolve("no-placement.yaml");

bridged/src/test/java/dev/ltms/bridged/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.CopyOnWriteArrayList;
 
 /**
  * Recording fake {@link HerdrClient} for unit/acceptance tests. Returns canned frames
@@ -22,7 +23,13 @@ public final class FakeHerdr implements HerdrClient {
     public static final long WORKER_PID = 4242;
 
     private final ObjectMapper mapper = new ObjectMapper();
-    public final List<Call> calls = new ArrayList<>();
+    /**
+     * Thread-safe on purpose. Background loops — {@link dev.ltms.bridged.msg.ReplyPushLoop} and the
+     * lead heartbeat — call this fake from their own scheduler threads while a test polls
+     * {@link #called} from the test thread. A plain {@code ArrayList} threw
+     * {@code ConcurrentModificationException} out of {@code called()} when a nudge landed mid-stream.
+     */
+    public final List<Call> calls = new CopyOnWriteArrayList<>();
     private boolean healthy = true;
     private final List<String> extraWorkspaces = new ArrayList<>();
     private final List<String> extraAgents = new ArrayList<>();

bridged/src/test/java/dev/ltms/bridged/mcp/BridgeMcpTest.java

@@ -16,12 +16,15 @@ import dev.ltms.bridged.inject.MemberPresence;
 import dev.ltms.bridged.session.MemberSession;
 import dev.ltms.bridged.session.WorktreeRequest;
 import dev.ltms.bridged.member.ClaudeCodeLauncher;
+import dev.ltms.bridged.member.CompositePeerLauncher;
 import dev.ltms.bridged.placement.BackendQuarantine;
+import dev.ltms.bridged.placement.PlacementPolicies;
 import io.modelcontextprotocol.spec.McpSchema;
 import dev.ltms.bridged.msg.InMemoryReplyInbox;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 
+import java.util.List;
 import java.util.Map;
 import java.util.Set;
 import java.util.concurrent.CompletableFuture;
@@ -423,6 +426,35 @@ class BridgeMcpTest {
         assertTrue(textOf(res).contains("unknown worker profile"), textOf(res));
     }
 
+    /**
+     * CB-599: a profile at its {@code maxLoad} cap must surface a readable reason on the MCP
+     * surface too, not merely flip {@code isError} with an opaque or absent message.
+     */
+    @Test
+    void spawnAtMaxLoadSurfacesTheCapacityReason() {
+        FakeHerdr h = new FakeHerdr();
+        BridgedConfig.Profile wcfg = new BridgedConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "BRIDGED_WORKER_TOKEN", null,
+                "tab", "bridged-workers", "worker: {profile} #{n}", null,
+                null, null, null, null, null, null, null, 0, null, null, null);
+        Map<String, BridgedConfig.Profile> profiles = Map.of(wcfg.profile(), wcfg);
+        ClaudeCodeLauncher delegate = new ClaudeCodeLauncher(
+                new AgentControl(h), new WorkspaceControl(h), new SubscriptionGuard(Set.of("gx00.gw")),
+                profiles, wcfg.profile(), k -> "BRIDGED_WORKER_TOKEN".equals(k) ? "tok" : null);
+        CompositePeerLauncher composite = new CompositePeerLauncher(
+                List.of(delegate), wcfg.profile(), profiles, PlacementPolicies.fixed(), _ -> 0);
+        SessionManager sm = new SessionManager(composite);
+
+        McpSchema.CallToolResult res = BridgeMcp.spawn(sm, "ltms-local");
+
+        assertTrue(res.isError());
+        String text = textOf(res);
+        assertTrue(text.contains("no capacity"), "surfaces a capacity reason, not a bare error: " + text);
+        assertTrue(text.contains("ltms-local"), "names the profile: " + text);
+        assertTrue(text.contains("maxLoad"), "explains the refusal: " + text);
+        assertFalse(h.called("agent.start"), "at cap, the spawn is refused before any herdr call");
+    }
+
     @Test
     void spawnPassesTheRequestedCwdToTheWorker() {
         FakeHerdr h = new FakeHerdr();

bridged/src/test/java/dev/ltms/bridged/msg/AmqpReplyInboxRecoveryRaceTest.java

@@ -35,9 +35,12 @@ import static org.junit.jupiter.api.Assertions.assertTrue;
  */
 class AmqpReplyInboxRecoveryRaceTest {
 
-    /** Large enough that the (unfixed) unsynchronized sweep's iteration is a real, observable window
-     *  a concurrently-started publish can land in — not just a best case, single-entry sprint. */
-    private static final int STALE_PUBLISHES = 100_000;
+    /** Large enough that thousands of entries are still unprocessed by the time the very first one
+     *  is observed as failed (see {@code sweepIsHoldingTheLock} below) — that gap is what makes the
+     *  head start deterministic instead of a coin flip. 100,000 gave the same guarantee but made the
+     *  test far more expensive than the guarantee needs; the ordering no longer depends on a timing
+     *  window sized to the full backlog; just to the tail of it. */
+    private static final int STALE_PUBLISHES = 2_000;
 
     @Test
     @Timeout(30)
@@ -58,6 +61,12 @@ class AmqpReplyInboxRecoveryRaceTest {
         // pendingByMsgId exactly like publishes whose confirm never arrived before a connection drop.
         // Virtual threads make this many concurrent blocking publish() calls cheap.
         CountDownLatch staleStarted = new CountDownLatch(STALE_PUBLISHES);
+        // Counted down by the FIRST stale publish thread to observe its own failure. That can only
+        // happen from inside failPendingPublishesOnRecovery() — nothing else in this test ever
+        // completes a stale Pending exceptionally (no nack/return is simulated for any "stale-*"
+        // msgId) — so seeing it fire is direct, observable proof the sweep is inside its loop, not a
+        // timing guess. It replaces the old fixed Thread.sleep(5) head start.
+        CountDownLatch sweepIsHoldingTheLock = new CountDownLatch(1);
         for (int i = 0; i < STALE_PUBLISHES; i++) {
             String msgId = "stale-" + i;
             Thread.ofVirtual().start(() -> {
@@ -65,7 +74,7 @@ class AmqpReplyInboxRecoveryRaceTest {
                 try {
                     inbox.publish("worker-stale", msgId, "x");
                 } catch (IllegalStateException expected) {
-                    // resolved (failed by the sweep) — that is exactly what this thread is here for
+                    sweepIsHoldingTheLock.countDown();
                 }
             });
         }
@@ -83,13 +92,28 @@ class AmqpReplyInboxRecoveryRaceTest {
             }
         }, "recovery-sweep");
         sweepThread.start();
-        // A short, deliberate head start: with STALE_PUBLISHES this large, the (unfixed) sweep's own
-        // iteration takes several milliseconds, so this guarantees the sweep has already begun —
-        // and, once guarded, is already holding publishChannelLock — before "fresh" attempts to
-        // register. Without this head start, "fresh" sometimes wins the race for the lock and
-        // registers before the sweep even starts, which is the accepted "already in flight when
-        // recovery fires" case (correctly failed either way) rather than the bug under test.
-        Thread.sleep(5);
+
+        // Deterministic head start: block until the sweep has actually failed one of the stale
+        // publishes. failPendingPublishesOnRecovery() (once guarded, as it is on main) holds
+        // publishChannelLock for its ENTIRE loop, not just per entry — so this failure proves the
+        // sweep is, at this instant, still holding that lock. With STALE_PUBLISHES this large, the
+        // remaining ~1,999 entries give an enormous margin between "first failure observed" and "sweep
+        // releases the lock": there is no window left for "fresh" to slip in before the sweep starts,
+        // or to win the lock ahead of it — see the case-2 note below. This also means Case 1 (the sweep
+        // is already inside its loop, holding the lock, when "fresh" tries to register) is now
+        // guaranteed by construction rather than merely likely under a fixed sleep.
+        assertTrue(sweepIsHoldingTheLock.await(20, TimeUnit.SECONDS),
+                "the sweep never failed a single stale publish — it may not have started");
+
+        // Case 2 ("fresh" wins publishChannelLock before the sweep even starts, so it genuinely
+        // published on the stale channel and the sweep correctly fails it) is impossible by
+        // construction in this test: freshThread.start() below is reached only after
+        // sweepIsHoldingTheLock has counted down, which can only happen once
+        // failPendingPublishesOnRecovery() is already running and has already failed a stale entry.
+        // There is no code path that lets "fresh" start before the sweep starts. That case is real
+        // and correct production behaviour (see AmqpReplyInbox#failPendingPublishesOnRecovery's
+        // javadoc), it is just not reachable from this deterministic ordering, so it does not need a
+        // separate assertion here.
 
         // This is the exact interleaving CB-528's follow-up describes: "the still-running recovery
         // sweep" racing a publish that registers while it is mid-flight.
@@ -109,8 +133,8 @@ class AmqpReplyInboxRecoveryRaceTest {
         // Simulate the broker's real confirm for "fresh" now that the sweep is done, so a correct
         // implementation's publish() returns normally instead of idling out CONFIRM_TIMEOUT_MS. Poll
         // for the registration rather than checking once: freshThread may still be contending for
-        // publishChannelLock (behind the 20,000 stale threads' own lock acquisitions) even though the
-        // sweep itself has already finished.
+        // publishChannelLock (behind the sweep's own hold on it, and possibly other stale threads
+        // still unwinding) even though the sweep itself has already finished.
         long deadline = System.nanoTime() + TimeUnit.SECONDS.toNanos(9);
         int idx = -1;
         while (idx < 0 && System.nanoTime() < deadline) {

bridged/src/test/java/dev/ltms/bridged/rest/BridgedAppTest.java

@@ -18,6 +18,8 @@ import dev.ltms.bridged.session.GitWorktrees;
 import dev.ltms.bridged.session.SessionManager;
 import dev.ltms.bridged.session.Worktrees;
 import dev.ltms.bridged.member.ClaudeCodeLauncher;
+import dev.ltms.bridged.member.CompositePeerLauncher;
+import dev.ltms.bridged.placement.PlacementPolicies;
 import io.javalin.Javalin;
 import org.junit.jupiter.api.AfterEach;
 import org.junit.jupiter.api.Test;
@@ -240,6 +242,43 @@ class BridgedAppTest {
         assertFalse(herdr.called("agent.start"), "an unknown profile must not spawn anything");
     }
 
+    /**
+     * CB-599: a profile at its {@code maxLoad} cap must not surface as a bare 500 — the caller
+     * needs a structured, readable reason, distinct from "unknown_profile".
+     */
+    @Test
+    void spawnAtMaxLoadIs503WithTheCapacityReasonNotABare500() throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        BridgedConfig.Profile wcfg = new BridgedConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "BRIDGED_WORKER_TOKEN", null,
+                "tab", "bridged-workers", "worker: {profile} #{n}", null,
+                null, null, null, null, null, null, null, 0, null, null, null);
+        Map<String, BridgedConfig.Profile> profiles = Map.of(wcfg.profile(), wcfg);
+        ClaudeCodeLauncher delegate = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr), new SubscriptionGuard(Set.of("gx00.gw")),
+                profiles, wcfg.profile(), k -> "BRIDGED_WORKER_TOKEN".equals(k) ? "tok-abc" : null);
+        CompositePeerLauncher workers = new CompositePeerLauncher(
+                List.of(delegate), wcfg.profile(), profiles, PlacementPolicies.fixed(), _ -> 0);
+        SessionManager sessions = new SessionManager(workers, new GitWorktrees());
+        this.presence = sessions.asPresence();
+        Injector injector = new Injector(new AgentControl(herdr));
+        InMemoryReplyInbox inbox = new InMemoryReplyInbox();
+        sessions.onAcquire(inbox::own);
+        MessageService messages = new MessageService(new AgentControl(herdr), injector, new Rendezvous(), inbox);
+        app = new BridgedApp(herdr, workers, sessions, messages, this.presence, null).build().start("127.0.0.1", 0);
+        int port = app.port();
+
+        HttpResponse<String> res = req(port, "POST", "/members?profile=ltms-local");
+
+        assertEquals(503, res.statusCode(), res.body());
+        JsonNode body = mapper.readTree(res.body());
+        assertEquals("no_capacity", body.get("error").asText());
+        String detail = body.get("detail").asText();
+        assertTrue(detail.contains("ltms-local"), "detail names the profile: " + detail);
+        assertTrue(detail.contains("maxLoad"), "detail explains the refusal: " + detail);
+        assertFalse(herdr.called("agent.start"), "at cap, the spawn is refused before any herdr call");
+    }
+
     @Test
     void spawnWorkerReusesExistingWorkerSpace() throws Exception {
         // A space labelled "bridged-workers" already exists → no second workspace.create.

bridged/src/test/java/dev/ltms/bridged/session/WorktreeSessionManagerTest.java

@@ -422,6 +422,28 @@ class WorktreeSessionManagerTest {
                 "acceptance criterion 6: a failed ticket's detail must carry the snapshot ref");
     }
 
+    @Test
+    void releaseNotifiesTheListenerWithTheAgentSessionId() {
+        // CB-584 (issue #65 criterion 5): a failed ticket's detail must also name the agent
+        // session, alongside worktree/branch/snapshot, so a lead can resume the conversation
+        // rather than only re-dispatch a fresh member onto the same files.
+        FakeHerdr herdr = new FakeHerdr();
+        FakeWorktrees worktrees = new FakeWorktrees().withRepoRoot("/repo").withPrefix("/wt")
+                .withDirty(true);
+        SessionManager sessions = new SessionManager(workerService(herdr), worktrees);
+        java.util.List<SessionManager.ReleaseDetail> released = new java.util.concurrent.CopyOnWriteArrayList<>();
+        sessions.onRelease(released::add);
+        MemberSession s = sessions.acquire("ltms-local", MemberRole.DEV, null, "/caller/proj", null,
+                new WorktreeRequest("cb-584-e", null), "cb-584-session", null);
+
+        assertNotNull(s.agentSessionId(), "a named session must mint an agent session id to assert on");
+        sessions.release(s.paneId());
+
+        assertEquals(1, released.size());
+        SessionManager.ReleaseDetail detail = released.getFirst();
+        assertEquals(s.agentSessionId(), detail.agentSessionId());
+    }
+
     @Test
     void aFailingSnapshotStillPreservesTheWorktreeStopsThePaneAndNotifies() {
         FakeHerdr herdr = new FakeHerdr();

deploy/dev.ltms.bridged.plist

@@ -82,8 +82,25 @@
     <key>RunAtLoad</key>
     <true/>
 
-    <!-- Restart on crash, but not in a tight loop if the config is bad (bridged fails fast on a
-         non-loopback bind without token auth — that is a config error, not a transient one). -->
+    <!--
+      CB-600 — read this before assuming ThrottleInterval bounds anything. It paces restarts to at
+      most one per 10s; it does NOT cap how many times launchd retries. If bridged fails fast on
+      every start — a bad bridged.yaml, for example auth.mode: token with the token env var unset,
+      which throws in main() before the daemon ever binds a port — launchd restarts it forever,
+      once every 10s, until a human intervenes. LaunchAgents have no "give up after N attempts"
+      primitive, so this is not something a config change here can fix.
+
+      That loop stops only two ways: (1) `launchctl unload -w ~/Library/LaunchAgents/dev.ltms.bridged.plist`,
+      or (2) the underlying cause gets fixed, so the process starts successfully and stays up (no
+      more exits to restart). scripts/redeploy-bridged.sh does not add a third way — it does not
+      make bridged self-disable on a config error, on purpose: a fail-fast exit path that
+      sometimes decides "this is unrecoverable, stop trying" is one more thing that can misfire,
+      and a wrongly self-disabled daemon needs the exact same manual `launchctl load -w` recovery
+      this comment already names — so it buys nothing an operator watching for the crash loop
+      doesn't already have, at the cost of a new way to be silently down. Watch for it with
+      `launchctl list dev.ltms.bridged` (a high restart count) or by tailing bridged.out for the
+      same startup error repeating every ~10s.
+    -->
     <key>KeepAlive</key>
     <dict>
         <key>SuccessfulExit</key>

scripts/redeploy-bridged.sh

@@ -79,6 +79,53 @@ running_pid() { pgrep -f "$PATTERN" || true; }
 launchd_installed() { [ -f "$LAUNCHD_PLIST" ]; }
 launchd_loaded() { launchctl list "$LAUNCHD_LABEL" >/dev/null 2>&1; }
 
+# CB-600: the script computes its own log path from where it sits on disk (REPO, above); the
+# plist hard-codes an absolute StandardOutPath. Nothing forced the two to agree — if this script
+# were ever run from a checkout other than the one the loaded plist names, launchd would start and
+# log the daemon correctly, while every check below (the fresh "bridged listening" line, the
+# ERROR-count scan) would read a different, empty or stale file and the script would report a
+# clean restart while the daemon crash-loops. Pure and side-effect-free besides `die`/`ok` — reads
+# the two paths, resolves them, compares — so it never touches launchd or the daemon and can be
+# exercised by sourcing this script (see the SOURCED guard below) without installing the agent.
+check_log_path_matches_plist() {
+  local script_out="$1" plist_path="$2"
+  local plist_out resolved_out resolved_plist_out
+  # Checked by exit status, not by emptiness: on a missing file/key PlistBuddy exits nonzero but
+  # still writes a message ("File Doesn't Exist, Will Create: ...") that command substitution
+  # would happily capture as if it were the real value — testing only `-z` missed that case.
+  if ! plist_out="$(/usr/libexec/PlistBuddy -c 'Print :StandardOutPath' "$plist_path" 2>/dev/null)" \
+      || [ -z "$plist_out" ]; then
+    die "launchd agent is loaded but PlistBuddy could not read StandardOutPath from
+        $plist_path
+      — cannot verify the daemon logs where this script is about to look. Fix the plist before
+      redeploying supervised."
+  fi
+  resolved_out="$(cd "$(dirname "$script_out")" 2>/dev/null && pwd -P)/$(basename "$script_out")" || true
+  resolved_plist_out="$(cd "$(dirname "$plist_out")" 2>/dev/null && pwd -P)/$(basename "$plist_out")" || true
+  if [ -z "$resolved_out" ] || [ -z "$resolved_plist_out" ] || [ "$resolved_out" != "$resolved_plist_out" ]; then
+    die "log path mismatch — this script reads
+        $script_out  (resolved: ${resolved_out:-<directory does not exist>})
+      but the loaded plist's StandardOutPath is
+        $plist_out  (resolved: ${resolved_plist_out:-<directory does not exist>})
+      Under supervision the daemon writes to the PLIST's path, not necessarily this script's — every
+      post-restart check below (the fresh 'bridged listening' line, the ERROR-count scan) would read
+      the wrong file and could report a clean restart while the daemon crash-loops. Fix the mismatch
+      (move this checkout to match the plist, or edit the plist's StandardOutPath/StandardErrorPath)
+      before redeploying supervised."
+  fi
+  ok "log path check: script and plist agree ($resolved_out)"
+}
+
+# CB-600: sourceable for testing. When this file is SOURCED (not executed) it stops here — nothing
+# below runs — so a test harness can `source` it to call check_log_path_matches_plist (or the
+# other pure helpers above) against a throwaway plist fixture without ever reaching the mutating
+# flow (build/stop/start) or touching the real daemon or launchd. On a normal `./redeploy-bridged.sh`
+# invocation `(return 0 2>/dev/null)` fails (return is illegal at top level of an executed script),
+# so this whole block is a no-op and every line below still runs exactly as before.
+if (return 0 2>/dev/null); then
+  return 0
+fi
+
 # ---------------------------------------------------------------- report state
 
 say "current state"
@@ -103,6 +150,9 @@ SUPERVISED=0
 if launchd_loaded; then
   SUPERVISED=1
   ok "launchd agent loaded ($LAUNCHD_LABEL) — launchd supervises this daemon"
+  # CB-600: fail loudly here, before ANY other check runs, if this script and the loaded plist
+  # would read different log files — every check after this point is worthless otherwise.
+  check_log_path_matches_plist "$OUT" "$LAUNCHD_PLIST"
 else
   warn "launchd agent not loaded — this script is the only thing that will restart the daemon."
 fi
@@ -222,7 +272,22 @@ say "start"
 if [ "$SUPERVISED" = 1 ]; then
   echo "   supervision is ON: using 'launchctl load' so launchd starts and keeps supervising this"
   echo "   process, instead of a manual nohup that launchd would know nothing about."
-  launchctl load -w "$LAUNCHD_PLIST" || die "launchctl load failed"
+  # CB-600: 'launchctl unload -w' above already persisted Disabled=true for this label. A load -w
+  # that succeeds clears it; a load -w that FAILS leaves the agent both stopped and disabled — worse
+  # than before this script ran, because a later reboot or login will not bring it back either. One
+  # retry covers a transient race (e.g. launchd not yet fully done deregistering); if it still fails,
+  # die with the exact recovery command rather than a bare "failed".
+  if ! launchctl load -w "$LAUNCHD_PLIST" 2>/dev/null; then
+    warn "launchctl load failed on the first attempt — retrying once after a short pause"
+    sleep 2
+    launchctl load -w "$LAUNCHD_PLIST" || die "launchctl load failed twice.
+        The agent is now STOPPED and DISABLED — it will NOT come back on its own, not even after a
+        reboot or login, because 'launchctl unload -w' above persisted Disabled=true and load -w
+        never got the chance to clear it. Recover with:
+            launchctl load -w \"$LAUNCHD_PLIST\"
+        If that still fails, check 'launchctl list $LAUNCHD_LABEL', validate the plist with
+        'plutil -lint \"$LAUNCHD_PLIST\"', and check $OUT before assuming a retry will succeed."
+  fi
 else
   # Absolute jar path so `ps` names which checkout is running.
   ( cd "$BRIDGED" && zsh -lc "nohup java -jar '$JAR' >> bridged.out 2>&1 &" )