Merge PR #699: fleetd #669 Unit C — the COLLABORATOR role and its authorization row

Adds Role.COLLABORATOR, Principal.collaborator(), and the Authz.permits grant:
READ/METRICS open, REPLY/ASK via ownership, SEND limited to a configured lead
or collaborator via a new classifier parameter, everything else denied.
isSpawnedMember() stays WORKER || ARCHITECT. fleet_whoami reports role and
collaborator name with no leader key.

Per a same-day ticket correction, the existing 3-argument Authz.permits is
kept (fail-closed default via the new NO_KNOWN_LEAD_OR_COLLABORATOR
classifier) rather than deleted, so the 47 pre-existing call sites in
AuthzTest/CallerResolverTest are untouched; both production gates
(FleetMcp#denyFor, FleetApp#allow via the new permitsFor seam) call the
4-argument form explicitly with the shared constant.

Mutation-verified: removing the classifier conjunct from the SEND arm kills
exactly 3 tests (AuthzTest, FleetMcpAuthzTest, FleetAppAuthTest), one per
gate, no cascade.

Full mvn clean install at this commit: 1974 tests, 0 failures (baseline at
f0ff252 was 1964; +10 are the new collaborator-matrix tests). Independently
counted from target/surefire-reports/*.xml after rm -rf: 173 report files,
aggregate tests=1974 failures=0 errors=0.

.claude/skills/fleets-status/SKILL.md

@@ -59,7 +59,7 @@ as `matches HEAD`, `drift`, or `unknown`; do not turn an unclear timestamp into
 Report the process identifier (PID) and uptime too:
 
 ```bash
-PIDS="$(pgrep -f 'target/fleetd.jar' || true)"
+PIDS="$(pgrep -f 'fleetd.jar' || true)"
 if [ -z "$PIDS" ]; then
   printf '%s\n' 'fleetd: not running'
 else

.claude/skills/handover/SKILL.md

@@ -164,6 +164,13 @@ fails.
 - **`accepted` does not mean your pane has been cleared.** It means every gate passed and the roll
   is scheduled to run once your current turn ends. Say your goodbye in the same turn — you will not
   get another one.
+- **If you are still running after that turn, the roll did not happen.** A roll that works clears
+  you, so surviving your own goodbye is itself the signal that it refused. Check with
+  `fleet_handover{action: "status", token}`, using the token you confirmed. `TURN_NEVER_SETTLED`
+  means your turn ran past `leadRollover.turnSettleSeconds` and **no `/clear` was ever sent**: your
+  context is intact and nothing was lost. Open a fresh request and retry. Never assume the roll
+  succeeded because `confirm` answered `accepted` — by the time it refuses, there is no caller left
+  to tell, so this check is the only thing that closes that gap.
 - **There is no terminal or session parameter, on purpose.** The pane is always your own, resolved
   from your connection, so you can only ever roll yourself.
 - **`operatorConfirmed` is your report of what a human told you.** Do not pass `true` because you

.claude/skills/redeploy-fleetd/SKILL.md

@@ -22,13 +22,22 @@ scripts/redeploy-fleetd.sh --yes     # skip the drain prompt (fleet already chec
 scripts/redeploy-fleetd.sh --no-build # restart the jar already on disk
 ```
 
-`--no-build` skips the build and restarts whatever jar is at `fleetd/target/fleetd.jar`. Use it only
+`--no-build` skips the build and restarts whatever jar is at `fleetd/run/fleetd.jar` — the runtime
-when you just built and nothing changed since. It gives up the protection in the next paragraph: no
+path, not Maven's output path. Use it only when you just built and nothing changed since. It gives
-build runs, so a stale or missing jar is not caught early. The script still checks the file is there
+up the protection in the next paragraph: no build runs, so a stale or missing jar is not caught
-and dies with `no jar at … — run without --no-build` if it is not, but it cannot tell you the jar is
+early. The script still checks the file is there and dies with `no jar at … — run without
-old. A `mvn clean` in the tree deletes that jar while the daemon keeps running on it, and nothing
+--no-build` if it is not, but it cannot tell you the jar is old.
-degrades until the next restart. Run `--check` first: it prints the jar's hash and its modification
+
-time, so you can see for yourself whether the jar is missing or older than the code you mean to ship.
+The daemon runs from `fleetd/run/fleetd.jar`, not from `fleetd/target/fleetd.jar` where Maven
+writes its output (fleetd #664). That split is what makes a bare `mvn install`/`mvn clean` in the
+main clone harmless now: neither can reach the file the running daemon holds open, because that
+file no longer lives under `target/` at all. Verify a merge by building in a throwaway git
+worktree anyway — a build still produces nothing the fleet runs until this script's own `mv` of
+`target/fleetd.jar` onto `run/fleetd.jar`, performed only after the old daemon is confirmed gone.
+Let only `scripts/redeploy-fleetd.sh` touch `fleetd/run/fleetd.jar`. Run `--check` first: it prints
+the BUILT jar (`target/fleetd.jar`) and the RUNNING jar (`run/fleetd.jar`) as two separately
+labelled hash-and-mtime facts, so a mismatch between them — a build sitting unswapped, or a stale
+runtime jar — is visible before you decide anything.
 
 It builds before it stops anything, so a failed build never leaves the fleet down; it waits for the
 old process to exit rather than assuming; it polls `/healthz`; and it anchors its log checks to a

CLAUDE.md

@@ -323,6 +323,13 @@ must obey belongs in the charter, not here.
   reference**, with the intent→tool table above as the short form. `McpContractDocTest` fails if
   that page names a `fleet_*` tool the server does not register. The flows are kept out of this
   file because this file loads into every session's context.
+- **The daemon runs from `fleetd/run/fleetd.jar`, not `fleetd/target/fleetd.jar`** (fleetd #664).
+  Maven's own output still lands at `fleetd/target/fleetd.jar` — that part of the build is
+  unchanged — but the running daemon never has that file open, so a bare `mvn install`/`mvn clean`
+  in the main clone no longer corrupts anything a live process is reading. Verify merges in a
+  throwaway git worktree anyway: a build in the main clone still ships nothing until
+  `scripts/redeploy-fleetd.sh` moves it into place with its own atomic `mv`, performed only after
+  the old daemon is confirmed gone. Let only that script touch `fleetd/run/fleetd.jar`.
 
 ### Redeploying the daemon — the lead may do this (primary only)
 

deploy/dev.ltms.fleetd.plist

@@ -47,7 +47,7 @@
         <string>/Users/dai.ha/LTMS/claude-bridge/scripts/fleetd-launchd-wrapper.sh</string>
         <string>/Users/dai.ha/Softwares/jdks/jdk-25.0.3.jdk/Contents/Home/bin/java</string>
         <string>-jar</string>
-        <string>/Users/dai.ha/LTMS/claude-bridge/fleetd/target/fleetd.jar</string>
+        <string>/Users/dai.ha/LTMS/claude-bridge/fleetd/run/fleetd.jar</string>
         <string>fleetd.yaml</string>
     </array>
 

deploy/fleetd.service

@@ -50,7 +50,7 @@ WorkingDirectory=%h/LTMS/fleetd/fleetd
 # and looks healthy, and the failure appears hours later as a member that cannot open a pull
 # request. exec keeps it one process, so systemd tracks the right PID.
 # This also avoids a SECOND copy of the secrets in a systemd drop-in: one source of truth.
-ExecStart=/bin/zsh -lc "exec java -jar target/fleetd.jar fleetd.yaml"
+ExecStart=/bin/zsh -lc "exec java -jar run/fleetd.jar fleetd.yaml"
 
 # PrivateTmp MUST stay false -- see herdr.service. fleetd creates the member ZDOTDIR scrub dir and
 # the opencode config dir under java.io.tmpdir, and the member pane (a herdr child, a different

fleetd/.gitignore

@@ -2,6 +2,10 @@
 target/
 dependency-reduced-pom.xml
 
+# The daemon's runtime jar (fleetd #664). scripts/redeploy-fleetd.sh moves the built jar here
+# with a same-filesystem rename; this is never Maven's output path and never belongs in git.
+run/
+
 # Local runtime config (copy from fleetd.example.yaml). Both names are ignored: fleetd.yaml is
 # the current name, and bridged.yaml is the legacy name Fleetd still falls back to.
 fleetd.yaml

fleetd/fleetd.example.yaml

@@ -669,6 +669,15 @@ fleet:
   #     kind: opencode
   #     model: openai/gpt-5.6-terra
 
+  # A collaborator tab, keyed by name (fleetd #669). This block is parsed and validated today;
+  # nothing yet recognises or addresses the tab it names. Recognise-only, like a profile-less
+  # `leaders:` entry above: there is no `profile:`, no `instances:` and no `kind:`. `tab:` is
+  # REQUIRED and is the only field identity depends on, matched case-insensitively — the same
+  # GET-THE-VALUE-RIGHT warning above the `leaders:` block applies here too.
+  # collaborators:
+  #   reviewer-alex:
+  #     tab: "collab: alex"
+
   # architects:
   #   architect-1:
   #     profile: opus                # a strong model, on the operator's subscription

fleetd/pom.xml

@@ -189,8 +189,10 @@
 
     <build>
         <!-- CB-634: the cutover renamed the module dir (bridged/ -> fleetd/), the jar, and the
-         launchd plist together. The installed plist names fleetd/target/fleetd.jar and
+         launchd plist together. fleetd #664: the installed plist and the systemd unit now name
-         KeepAlive is armed, so this name, the plist, and the wrapper must move as one. -->
+         fleetd/run/fleetd.jar, not this plugin's own output path — see
+         scripts/redeploy-fleetd.sh for the mv that gets a build from here to there. KeepAlive is
+         armed, so this name, the plist, and the wrapper must still move as one. -->
         <finalName>fleetd</finalName>
         <plugins>
             <plugin>

fleetd/src/main/java/dev/ltms/fleet/auth/Authz.java

@@ -1,5 +1,7 @@
 package dev.ltms.fleet.auth;
 
+import java.util.function.Predicate;
+
 /**
  * The authorization table (CB-505), stated once and enforced on both entry paths.
  *
@@ -20,16 +22,22 @@ public final class Authz {
         SPAWN,
         /** Tear a worker peer down. */
         STOP,
-        /** Deliver a turn to a session (or answer a worker's question). */
+        /** Deliver a turn to a local session, addressed by {@code sessionId}. */
         SEND,
+        /** Resolve a worker's blocked question and resume its turn, addressed by {@code turnId}. */
+        ANSWER,
+        /** Address a peer lead on another daemon over the coordination broker, by {@code coordId}. */
+        COORD_SEND,
         /** A worker's terminal reply for its own turn. */
         REPLY,
         /** A worker's mid-turn question to the primary. */
         ASK,
         /** Collect held replies from a session's inbox. */
         DRAIN,
-        /** Read-only observation: status, roster, profiles, task polling. */
+        /** Read-only roster, profile, and identity observation: no ticket, task, or turn state. */
         READ,
+        /** Poll a ticket, or read a session's status. */
+        TASK_READ,
         /**
          * Read (never ack) this daemon's own held lead-to-lead coordination mail (fleetd #421).
          *
@@ -54,43 +62,92 @@ public final class Authz {
     }
 
     /**
-     * Whether {@code caller} may perform {@code action} against {@code targetSession}.
+     * Stands in for the terminal-to-tab registry a collaborator's {@code SEND} is checked
-     *
+     * against, until one exists: answers no for every target, so a collaborator reaches nothing
-     * @param targetSession the session id in the request path; only consulted for the worker-scoped
+     * today. Both production gates ({@code FleetMcp#denyFor}, {@code FleetApp#allow}) pass this
-     *                      actions ({@code REPLY}, {@code ASK}), ignored otherwise, may be
+     * exact instance, so the classifier is defined once and replacing it is a one-line change in
-     *                      {@code null}
+     * each.
+     */
+    public static final Predicate<String> NO_KNOWN_LEAD_OR_COLLABORATOR = target -> false;
+
+    /**
+     * Convenience form for a caller with no classifier to supply. Fails closed: a collaborator's
+     * {@code SEND} is refused, as if no terminal were a configured lead or collaborator — the
+     * same decision {@link #NO_KNOWN_LEAD_OR_COLLABORATOR} gives explicitly. Every other action's
+     * result is identical to the four-argument form's, since none of them consult the classifier.
      */
     public static boolean permits(Principal caller, Action action, String targetSession) {
+        return permits(caller, action, targetSession, NO_KNOWN_LEAD_OR_COLLABORATOR);
+    }
+
+    /**
+     * Whether {@code caller} may perform {@code action} against {@code targetSession}.
+     *
+     * @param targetSession           the session id in the request path; only consulted for the
+     *                                worker-scoped actions ({@code REPLY}, {@code ASK}) and for a
+     *                                collaborator's {@code SEND}, ignored otherwise, may be
+     *                                {@code null}
+     * @param knownLeadOrCollaborator whether a terminal is a configured lead or collaborator —
+     *                                consulted only for a collaborator's {@code SEND}, to confine
+     *                                it to another named peer and never a spawned member's
+     *                                terminal
+     */
+    public static boolean permits(Principal caller, Action action, String targetSession,
+                                   Predicate<String> knownLeadOrCollaborator) {
         if (caller == null || caller.isAnonymous()) {
             return false; // authenticated as nothing ⇒ authorized for nothing
         }
         return switch (action) {
-            // Fleet lifecycle is the primary's alone — spawn, stop, drain. An architect
+            // Fleet lifecycle is the primary's alone — spawn, stop, drain. An architect and a
-            // deliberately does NOT get these (CB-548), so it cannot tear down or stand up workers
+            // collaborator deliberately do NOT get these, so neither can tear down or stand up
-            // even though it coordinates them; and a worker driving any of these would be a worker
+            // workers even though one of them coordinates them; and a worker driving any of these
-            // escalating into the orchestrator role.
+            // would be a worker escalating into the orchestrator role.
             case SPAWN, STOP, DRAIN, HANDOVER -> caller.isPrimary();
 
-            // Delivering a turn is open to the primary and the architect: an architect delegates
+            // Delivering a turn to a local session is open to the primary, the architect, and a
-            // to workers (that is the role's point) but still has no lifecycle rights. A worker is
+            // collaborator whose target is itself a configured lead or collaborator: the architect
-            // excluded — sending would be it escalating.
+            // delegates to workers (that is the role's point); a collaborator may reach only
-            case SEND -> caller.isPrimary() || caller.isArchitect();
+            // another named peer, never a spawned member's terminal. A worker is excluded —
+            // sending would be it escalating.
+            case SEND -> caller.isPrimary() || caller.isArchitect()
+                    || (caller.isCollaborator() && knownLeadOrCollaborator.test(targetSession));
+
+            // Resolving a worker's blocked question is part of delegating to it, open to the same
+            // two roles that may stand up that delegation in the first place. Not a collaborator:
+            // resuming another session's turn is lifecycle-adjacent, not peer messaging.
+            case ANSWER -> caller.isPrimary() || caller.isArchitect();
+
+            // Leaves the daemon over the coordination broker rather than addressing a local
+            // session, open to the same two roles as ANSWER. Not a collaborator: it is a
+            // local-tab peer with no cross-host route.
+            case COORD_SEND -> caller.isPrimary() || caller.isArchitect();
 
             // The load-bearing rule: a caller acts only as the pane it occupies. CB-532 widened who
             // that can be — a lead answering another lead is replying for its OWN terminal, which
             // this already permits — while the rule itself is unchanged, and is what stops anyone
-            // forging a reply for a rendezvous someone else is waiting on. An architect's own pane
+            // forging a reply for a rendezvous someone else is waiting on. An architect's or a
-            // passes through the same check, so it can answer a funnel that delegated to it. An
+            // collaborator's own pane passes through the same check, so each can answer a funnel
-            // unnamed primary (token/loopback, no pane) owns nothing and is still excluded.
+            // that delegated to it. An unnamed primary (token/loopback, no pane) owns nothing and
+            // is still excluded.
             case REPLY, ASK -> caller.ownsSession(targetSession);
 
-            // Observation is open to every authenticated role: a worker legitimately polls its own
+            // READ is roster, profile, and identity observation — fleet_list, fleet_profiles, and
-            // status, and the roster carries no secrets.
+            // fleet_whoami — and carries no secrets: no ticket reply, no pending question, and no
-            case READ, METRICS -> caller.isPrimary() || caller.isWorker() || caller.isArchitect();
+            // other session's turn state. Those live under TASK_READ. METRICS is the separate
+            // Prometheus scrape. Both are open to every authenticated role, including a
+            // collaborator.
+            case READ, METRICS -> caller.isPrimary() || caller.isWorker() || caller.isArchitect()
+                    || caller.isCollaborator();
+
+            // Ticket polling and session status, open to every role READ is open to except a
+            // collaborator: ticket ids are a sequential counter with no owner check, so a holder
+            // could walk every ticket and read another session's delegation reply.
+            case TASK_READ -> caller.isPrimary() || caller.isWorker() || caller.isArchitect();
 
             // fleetd #421: reading held lead-to-lead mail is the primary's alone. An architect
             // holds READ today (CB-548), so "not primary" must mean not-architect here too — this
-            // is coordination between leads, not observation of the roster.
+            // is coordination between leads, not observation of the roster. The same reasoning
+            // excludes a collaborator.
             case COORD_READ -> caller.isPrimary();
         };
     }

fleetd/src/main/java/dev/ltms/fleet/auth/Principal.java

@@ -11,7 +11,9 @@ package dev.ltms.fleet.auth;
  * @param pid      the connecting process id, or {@code -1} when not resolvable (audit context)
  * @param name     for a lead resolved from the CB-530 {@code leaders:} registry, which lead it is;
  *                 for an architect resolved from the CB-548 {@code architects:} registry, which
- *                 slot it occupies; {@code null} for every other caller, including an unnamed primary
+ *                 slot it occupies; for a collaborator resolved from the {@code collaborators:}
+ *                 registry, which collaborator it is; {@code null} for every other caller,
+ *                 including an unnamed primary
  */
 public record Principal(Role role, String terminal, long pid, String name) {
 
@@ -73,6 +75,19 @@ public record Principal(Role role, String terminal, long pid, String name) {
         return new Principal(Role.ARCHITECT, terminal, pid, slotName);
     }
 
+    /**
+     * A collaborator: a human-opened tab recognised by its exact label in the
+     * {@code collaborators:} registry.
+     *
+     * <p>Carries {@link Role#COLLABORATOR}. {@code name} is reporting only — it lets
+     * {@code fleet_whoami} say which collaborator is asking. Identity is the {@code terminal}:
+     * like a worker's it comes from the connection, so {@code ownsSession} works exactly as it
+     * does for a worker — a collaborator acts as its own pane and no other.
+     */
+    public static Principal collaborator(String name, String terminal, long pid) {
+        return new Principal(Role.COLLABORATOR, terminal, pid, name);
+    }
+
     public boolean isPrimary() {
         return role == Role.PRIMARY;
     }
@@ -81,6 +96,10 @@ public record Principal(Role role, String terminal, long pid, String name) {
         return role == Role.ARCHITECT;
     }
 
+    public boolean isCollaborator() {
+        return role == Role.COLLABORATOR;
+    }
+
     public boolean isWorker() {
         return role == Role.WORKER;
     }
@@ -120,6 +139,7 @@ public record Principal(Role role, String terminal, long pid, String name) {
         return switch (role) {
             case WORKER -> "worker:" + terminal;
             case ARCHITECT -> "architect:" + name;
+            case COLLABORATOR -> "collaborator:" + name;
             case PRIMARY -> name == null ? "primary" : "leader:" + name;
             case ANONYMOUS -> "anonymous";
         };

fleetd/src/main/java/dev/ltms/fleet/auth/Role.java

@@ -34,6 +34,17 @@ public enum Role {
      */
     ARCHITECT,
 
+    /**
+     * A config-declared, human-opened tab recognised by its exact label (the {@code
+     * fleet.collaborators.<name>.tab} registry). Never spawned — identity comes from the
+     * connection, never a request argument, exactly like {@link #WORKER} and {@link #ARCHITECT}.
+     * May {@code SEND} only to a configured lead or collaborator, {@code REPLY}/{@code ASK} only
+     * as its own pane, and {@code READ}/{@code METRICS}; may not {@code SPAWN}/{@code STOP}/
+     * {@code DRAIN}/{@code HANDOVER}, poll a ticket ({@code TASK_READ}), or reach the
+     * coordination broker ({@code COORD_SEND}/{@code COORD_READ}).
+     */
+    COLLABORATOR,
+
     /** Authenticated as nothing. Authorized for nothing but {@code /healthz}. */
     ANONYMOUS
 }

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

@@ -1129,11 +1129,8 @@ public record FleetConfig(
      *                            {@code tab} can never be discovered, launched or not
      * @param instances           how many of this lead should be live (default 1). The daemon
      *                            launches only the shortfall, so a restart adopts rather than doubles
-     * @param tabPrefix           no longer used to find a lead's tab — {@code tab} is matched
+     * @param tabPrefix           lead-tab naming convention checked against member labels. Lead
-     *                            exactly. Its only remaining job is the startup collision guard
+     *                            identity uses {@code tab}. Default {@code "lead:"}
-     *                            ({@link #validateLeadTabPrefixes()}), which still uses it to refuse
-     *                            a worker {@code tabLabel} template that could be misread as a lead.
-     *                            Default {@code "lead:"}
      * @param scanIntervalSeconds how long a tab scan is cached before herdr is asked again; also the
      *                            worst case before a newly-labelled tab is recognised. Default 10
      * @param kind                which agent runs there ({@code claude}, {@code opencode}, …)
@@ -1179,6 +1176,25 @@ public record FleetConfig(
         }
     }
 
+    /**
+     * A tab fleetd recognises as a collaborator, keyed by name (fleetd #669).
+     *
+     * <p>Recognise-only: there is no {@code profile}, no {@code instances} and no {@code kind}.
+     * Nothing here ever launches a pane.
+     *
+     * <p>{@code tabPrefix} is absent. Identity is matched on the exact {@code tab} alone.
+     *
+     * @param tab the exact tab label hosting this collaborator, matched case-insensitively; the
+     *            only field identity depends on. Required — an entry with no {@code tab} can
+     *            never be discovered.
+     */
+    @JsonIgnoreProperties(ignoreUnknown = true)
+    public record Collaborator(String tab) {
+        public Collaborator {
+            tab = (tab == null || tab.isBlank()) ? null : tab.strip();
+        }
+    }
+
     /**
      * One entry of a {@code fleet:} role pool — a role paired with the backend it runs on.
      *
@@ -1216,15 +1232,18 @@ public record FleetConfig(
      * is exactly compatible with that. The pool is also what replaced {@code defaultProfile:} — an
      * unqualified spawn names a role, and the role's pool supplies the candidates.
      *
-     * @param leaders     panes that orchestrate rather than are orchestrated, keyed by lead name
+     * @param leaders       panes that orchestrate rather than are orchestrated, keyed by lead name
-     * @param architects  profiles the {@code architect} role may run on
+     * @param architects    profiles the {@code architect} role may run on
-     * @param developers  profiles the {@code dev} role may run on
+     * @param developers    profiles the {@code dev} role may run on
-     * @param hunters     profiles the {@code hunter} role may run on
+     * @param hunters       profiles the {@code hunter} role may run on
-     * @param reviewers   profiles the {@code reviewer} role may run on
+     * @param reviewers     profiles the {@code reviewer} role may run on
-     * @param charters    optional launch-charter text keyed by singular role wire name
+     * @param charters      optional launch-charter text keyed by singular role wire name
-     * @param tabLabel    template for a member tab's label; {@code {role}}, {@code {profile}},
+     * @param tabLabel      template for a member tab's label; {@code {role}}, {@code {profile}},
-     *                    {@code {model}} and {@code {n}} (a per role+profile counter) are
+     *                      {@code {model}} and {@code {n}} (a per role+profile counter) are
-     *                    substituted. Default {@link #DEFAULT_TAB_LABEL}
+     *                      substituted. Default {@link #DEFAULT_TAB_LABEL}
+     * @param collaborators tabs fleetd recognises as collaborators (fleetd #669), keyed by name.
+     *                      Recognise-only, exactly like a {@code profile}-less {@link Leader}:
+     *                      nothing here is ever auto-launched.
      */
     @JsonIgnoreProperties(ignoreUnknown = true)
     public record Fleet(Map<String, Leader> leaders,
@@ -1233,13 +1252,11 @@ public record FleetConfig(
                         Map<String, Slot> hunters,
                         Map<String, Slot> reviewers,
                         Map<String, String> charters,
-                        String tabLabel) {
+                        String tabLabel,
+                        Map<String, Collaborator> collaborators) {
 
         /**
-         * Role first, so the tab bar reads as the fleet and so the label shares a namespace with a
+         * Role first, so the tab bar identifies the member's fleet role.
-         * lead's {@code tabPrefix}. Because {@code {role}} comes from a closed enum, a generated
-         * member label can never begin with {@code "lead:"} — the clash that
-         * {@link #validateLeadTabPrefixes()} used to have to check for is unrepresentable here.
          */
         public static final String DEFAULT_TAB_LABEL = "{role}: {profile} #{n}";
 
@@ -1251,26 +1268,30 @@ public record FleetConfig(
             reviewers = unmodifiableOrEmpty(reviewers);
             charters = unmodifiableOrEmpty(charters);
             tabLabel = (tabLabel == null || tabLabel.isBlank()) ? DEFAULT_TAB_LABEL : tabLabel;
+            collaborators = unmodifiableOrEmpty(collaborators);
         }
 
         /**
-         * A fleet with no configured launch charters — the shape every deployment had before
+         * A fleet with no configured launch charters and no collaborators — the shape every
-         * CB-566, and what most tests want.
+         * deployment had before CB-566, and what most tests want.
          *
          * <p>Kept deliberately, even though an overload that drops a new field is normally the
          * shape to avoid. It is safe here because nothing <em>reads</em> a charter through a
          * constructor: the launcher reads {@code fleet.charters()} from the live config. Jackson
          * binds the canonical constructor, so this one cannot swallow an operator's YAML.
+         * {@code collaborators} is dropped the same way and for the same reason: no caller of
+         * this overload has ever needed to set it, so it defaults to empty here exactly as the
+         * canonical constructor would default an absent YAML key.
          */
         public Fleet(Map<String, Leader> leaders, Map<String, Slot> architects,
                      Map<String, Slot> developers, Map<String, Slot> reviewers,
                      Map<String, String> charters, String tabLabel) {
-            this(leaders, architects, developers, null, reviewers, charters, tabLabel);
+            this(leaders, architects, developers, null, reviewers, charters, tabLabel, null);
         }
 
         public Fleet(Map<String, Leader> leaders, Map<String, Slot> architects,
                      Map<String, Slot> developers, Map<String, Slot> reviewers, String tabLabel) {
-            this(leaders, architects, developers, null, reviewers, null, tabLabel);
+            this(leaders, architects, developers, null, reviewers, null, tabLabel, null);
         }
 
         /**
@@ -1399,11 +1420,13 @@ public record FleetConfig(
      * called FROM the calling lead's own turn, so its pane is still {@code WORKING} the instant
      * {@code confirm()} validates every gate and schedules the roll. {@code
      * dev.ltms.fleet.lead.LeadRollover}'s deferred continuation waits up to this many seconds for
-     * that SAME pane to report an injectable state again — i.e. for the calling turn to actually
+     * that SAME pane to report {@code IDLE} or {@code DONE} — i.e. for the calling turn to actually
-     * end — before it sends {@code /clear} at all. If that wait times out, no {@code /clear} is
+     * end — before it sends {@code /clear} at all. {@code BLOCKED} does not count: that is a live
+     * turn merely paused, not one that has finished. If that wait times out, no {@code /clear} is
      * ever sent: a lead that never goes idle is still doing real work, and clearing it would
      * destroy live context. This is a separate wait from {@code clearSettleSeconds} below, which
-     * bounds the SECOND wait, for the pane to re-settle AFTER {@code /clear} has already gone out.
+     * bounds the SECOND wait, for the pane to reach {@code IDLE} or {@code DONE} again AFTER
+     * {@code /clear} has already gone out.
      *
      * @param handoverPath required when this block is present — where the handover file a fresh
      *                     lead session reads must live. There is no sane non-null default for an
@@ -1422,12 +1445,13 @@ public record FleetConfig(
      * @param maxDocAgeSeconds default 3600 — refuse a handover file whose modified time is older
      *                     than this many seconds, so a stale leftover from an earlier rollover
      *                     attempt can never be mistaken for a fresh one.
-     * @param turnSettleSeconds default 20 — bound on how long the deferred roll waits for the
+     * @param turnSettleSeconds default 300 — bound on how long the deferred roll waits for the
-     *                     CALLING lead's own turn to end (its pane to report injectable again)
+     *                     CALLING lead's own turn to end (its pane to report {@code IDLE} or
-     *                     before sending {@code /clear} at all. See the paragraph above.
+     *                     {@code DONE}) before sending {@code /clear} at all. See the paragraph
+     *                     above.
      * @param clearSettleSeconds default 20 — bound on how long to wait for the lead's pane to
-     *                     report an injectable state again after {@code /clear} before giving up. A
+     *                     report {@code IDLE} or {@code DONE} again after {@code /clear} before
-     *                     roll that times out here never sends {@code bootstrapText}.
+     *                     giving up. A roll that times out here never sends {@code bootstrapText}.
      * @param bootstrapText default a sentence naming the RESOLVED handover path — sent to the
      *                     lead's pane once it settles after {@code /clear}, telling the fresh
      *                     session where to read the handover and carry on. Left {@code null} here
@@ -1444,7 +1468,7 @@ public record FleetConfig(
         public LeadRollover {
             requireOperatorConfirm = requireOperatorConfirm == null || requireOperatorConfirm;
             maxDocAgeSeconds = (maxDocAgeSeconds == null || maxDocAgeSeconds <= 0) ? 3600 : maxDocAgeSeconds;
-            turnSettleSeconds = (turnSettleSeconds == null || turnSettleSeconds <= 0) ? 20 : turnSettleSeconds;
+            turnSettleSeconds = (turnSettleSeconds == null || turnSettleSeconds <= 0) ? 300 : turnSettleSeconds;
             clearSettleSeconds = (clearSettleSeconds == null || clearSettleSeconds <= 0) ? 20 : clearSettleSeconds;
             bootstrapText = (bootstrapText == null || bootstrapText.isBlank()) ? null : bootstrapText;
         }
@@ -1913,7 +1937,7 @@ public record FleetConfig(
 
     /** The {@code fleet:} child blocks whose direct children are slot names. */
     private static final Set<String> FLEET_POOL_KEYS =
-            Set.of("leaders", "architects", "developers", "hunters", "reviewers");
+            Set.of("leaders", "architects", "developers", "hunters", "reviewers", "collaborators");
 
     /**
      * Reject a {@code fleet:} role pool whose slot names repeat (CB-548, re-homed by CB-557).
@@ -1923,7 +1947,7 @@ public record FleetConfig(
      * daemon would never know. Jackson's YAML parser does not fail on duplicate mapping keys by
      * default, so duplicates are caught here, at parse time, before the map is built.
      *
-     * <p>Only the five pools <em>directly under the top-level {@code fleet:}</em> are considered,
+     * <p>Only the six pools <em>directly under the top-level {@code fleet:}</em> are considered,
      * and only their direct child keys (the slot names). A nested field elsewhere, even one also
      * named {@code developers:}, is ignored, so parsing of the rest of the config is unaffected.
      *
@@ -2682,31 +2706,21 @@ public record FleetConfig(
     }
 
     /**
-     * Reject a lead-scan convention that a worker tab would also satisfy (CB-531).
+     * Reject a member tab-label template that could render as a configured lead or collaborator
+     * tab or match a lead-tab naming convention, and reject two {@code fleet.leaders} or
+     * {@code fleet.collaborators} entries — across either registry — that share one exact tab.
      *
-     * <p>The scan reads a tab label and concludes "a lead lives here". fleetd also <em>writes</em>
+     * <p>{@code fleet.collaborators} has no {@code tabPrefix}: identity is matched on the exact
-     * tab labels — every member gets one rendered into its tab. Choose a lead {@code tabPrefix} that
+     * {@code tab} alone, so only the exact-render check applies there, not the prefix check.
-     * a member template matches and the daemon starts labelling its own members as leads, promoting
-     * the entire fleet to {@link dev.ltms.fleet.auth.Role#PRIMARY} with no message and no diff.
-     * {@link #validatePanePlacementAgainstLeadTabs()} is the check that stops a pane-placed member
-     * from landing inside a lead's tab in the first place; this check is a second, independent
-     * guard that catches the hazard even when every profile places members correctly, by refusing
-     * a label that a scan would still misread as a lead.
      *
-     * <p>CB-557 shrank this check rather than removing it. The default template is
+     * @throws IllegalStateException when the fleet template or a profile {@code tabLabel} override
-     * {@code "{role}: {profile} #{n}"} and {@code {role}} comes from a closed enum, so a
+     *                               can render as a configured lead or collaborator tab or match a
-     * <em>generated</em> label can no longer collide by construction. What remains checkable is what
+     *                               lead-tab prefix, or when two entries — of either registry, or
-     * an operator still writes by hand: the {@code fleet.tabLabel} template and any per-profile
+     *                               one of each — carry the same exact {@code tab}
-     * {@code tabLabel} override.
+     *                               (case-insensitively)
-     *
-     * <p>Fatal rather than a warning, unlike {@link #warnUnknownTopLevelKeys}: an unknown key means
-     * a feature does nothing, while this means a feature does the opposite of what it says.
-     *
-     * @throws IllegalStateException when the fleet template or any profile's {@code tabLabel}
-     *                               override starts with a configured lead prefix
      */
     public void validateLeadTabPrefixes() {
-        if (fleet == null || fleet.leaders().isEmpty()) {
+        if (fleet == null) {
             return;
         }
         List<String> bad = new ArrayList<>();
@@ -2714,51 +2728,173 @@ public record FleetConfig(
             if (leader == null) {
                 return;
             }
+            String tab = leader.tab();
             String prefix = leader.tabPrefix();
-            // The fleet-wide template is checked once per prefix: it labels every member that has no
+            if (templateCanRenderAs(fleet.tabLabel(), tab)) {
-            // override, so one bad template promotes the entire fleet, not one profile.
+                bad.add("fleet.tabLabel=\"" + fleet.tabLabel() + "\" can render as the tab of "
-            if (startsWithIgnoreCase(fleet.tabLabel(), prefix)) {
+                        + "lead '" + leadName + "' (\"" + tab + "\")");
+            } else if (startsWithIgnoreCase(fleet.tabLabel(), prefix)) {
                 bad.add("fleet.tabLabel=\"" + fleet.tabLabel() + "\" starts with the tabPrefix of "
                         + "lead '" + leadName + "' (\"" + prefix + "\")");
             }
             profiles().entrySet().stream()
-                    .filter(e -> startsWithIgnoreCase(e.getValue().tabLabel(), prefix))
                     .map(Map.Entry::getKey)
                     .sorted()
-                    .forEach(p -> bad.add("profile '" + p + "' overrides tabLabel with \""
+                    .forEach(p -> {
-                            + profiles().get(p).tabLabel() + "\", which starts with the tabPrefix of "
+                        String label = profiles().get(p).tabLabel();
-                            + "lead '" + leadName + "' (\"" + prefix + "\")"));
+                        if (templateCanRenderAs(label, tab)) {
+                            bad.add("profile '" + p + "' overrides tabLabel with \"" + label
+                                    + "\", which can render as the tab of lead '" + leadName
+                                    + "' (\"" + tab + "\")");
+                        } else if (startsWithIgnoreCase(label, prefix)) {
+                            bad.add("profile '" + p + "' overrides tabLabel with \"" + label
+                                    + "\", which starts with the tabPrefix of lead '" + leadName
+                                    + "' (\"" + prefix + "\")");
+                        }
+                    });
         });
-        if (bad.isEmpty()) {
+        fleet.collaborators().forEach((collabName, collaborator) -> {
+            if (collaborator == null) {
+                return;
+            }
+            String tab = collaborator.tab();
+            if (templateCanRenderAs(fleet.tabLabel(), tab)) {
+                bad.add("fleet.tabLabel=\"" + fleet.tabLabel() + "\" can render as the tab of "
+                        + "collaborator '" + collabName + "' (\"" + tab + "\")");
+            }
+            profiles().entrySet().stream()
+                    .map(Map.Entry::getKey)
+                    .sorted()
+                    .forEach(p -> {
+                        String label = profiles().get(p).tabLabel();
+                        if (templateCanRenderAs(label, tab)) {
+                            bad.add("profile '" + p + "' overrides tabLabel with \"" + label
+                                    + "\", which can render as the tab of collaborator '"
+                                    + collabName + "' (\"" + tab + "\")");
+                        }
+                    });
+        });
+        if (!bad.isEmpty()) {
+            throw new IllegalStateException("refusing to start: " + String.join("; ", bad)
+                    + ". Every member labelled that way would be read back as a lead or "
+                    + "collaborator and granted that identity's authority. Change one of the two "
+                    + "so member tabs cannot be confused with a lead's or collaborator's tab.");
+        }
+
+        List<String> collisions = new ArrayList<>();
+        List<String> leadNames = fleet.leaders().keySet().stream().sorted().toList();
+        for (int i = 0; i < leadNames.size(); i++) {
+            String nameA = leadNames.get(i);
+            Leader a = fleet.leaders().get(nameA);
+            if (a == null || a.tab() == null || a.tab().isBlank()) {
+                continue;
+            }
+            for (int j = i + 1; j < leadNames.size(); j++) {
+                String nameB = leadNames.get(j);
+                Leader b = fleet.leaders().get(nameB);
+                if (b == null || b.tab() == null || b.tab().isBlank()) {
+                    continue;
+                }
+                if (a.tab().equalsIgnoreCase(b.tab())) {
+                    collisions.add("lead '" + nameA + "' and lead '" + nameB + "' both use tab \""
+                            + a.tab() + "\"");
+                }
+            }
+        }
+        List<String> collabNames = fleet.collaborators().keySet().stream().sorted().toList();
+        for (int i = 0; i < collabNames.size(); i++) {
+            String nameA = collabNames.get(i);
+            Collaborator a = fleet.collaborators().get(nameA);
+            if (a == null || a.tab() == null || a.tab().isBlank()) {
+                continue;
+            }
+            for (int j = i + 1; j < collabNames.size(); j++) {
+                String nameB = collabNames.get(j);
+                Collaborator b = fleet.collaborators().get(nameB);
+                if (b == null || b.tab() == null || b.tab().isBlank()) {
+                    continue;
+                }
+                if (a.tab().equalsIgnoreCase(b.tab())) {
+                    collisions.add("collaborator '" + nameA + "' and collaborator '" + nameB
+                            + "' both use tab \"" + a.tab() + "\"");
+                }
+            }
+        }
+        for (String leadName : leadNames) {
+            Leader lead = fleet.leaders().get(leadName);
+            if (lead == null || lead.tab() == null || lead.tab().isBlank()) {
+                continue;
+            }
+            for (String collabName : collabNames) {
+                Collaborator collaborator = fleet.collaborators().get(collabName);
+                if (collaborator == null || collaborator.tab() == null
+                        || collaborator.tab().isBlank()) {
+                    continue;
+                }
+                if (lead.tab().equalsIgnoreCase(collaborator.tab())) {
+                    collisions.add("lead '" + leadName + "' and collaborator '" + collabName
+                            + "' both use tab \"" + lead.tab() + "\"");
+                }
+            }
+        }
+        if (collisions.isEmpty()) {
             return;
         }
-        throw new IllegalStateException("refusing to start: " + String.join("; ", bad)
+        throw new IllegalStateException("refusing to start: " + String.join("; ", collisions)
-                + ". Every member labelled that way would be read back as a lead and granted "
+                + ". Tab identity is matched exactly, so only one of two entries sharing a tab can "
-                + "spawn/stop/send on the whole fleet. Change one of the two so member tabs and "
+                + "ever be found — the other is silently unreachable. Give each lead and "
-                + "lead tabs cannot be confused.");
+                + "collaborator its own exact tab.");
+    }
+
+    private static boolean templateCanRenderAs(String template, String tab) {
+        if (template == null || template.isBlank() || tab == null || tab.isBlank()) {
+            return false;
+        }
+        var placeholders = Pattern.compile("\\{(?:role|profile|model|n)}").matcher(template);
+        StringBuilder expression = new StringBuilder("^");
+        int literalStart = 0;
+        while (placeholders.find()) {
+            expression.append(Pattern.quote(template.substring(literalStart, placeholders.start())));
+            expression.append(".*");
+            literalStart = placeholders.end();
+        }
+        expression.append(Pattern.quote(template.substring(literalStart))).append("$");
+        return Pattern.compile(expression.toString(), Pattern.CASE_INSENSITIVE).matcher(tab).matches();
+    }
+
+    /** Case-insensitive prefix test that tolerates a null or blank label. */
+    private static boolean startsWithIgnoreCase(String label, String prefix) {
+        if (label == null || prefix == null || prefix.isBlank()) {
+            return false;
+        }
+        String stripped = label.strip();
+        return stripped.regionMatches(true, 0, prefix, 0, prefix.length());
     }
 
     /**
      * Reject a profile that places its members by {@code "pane"} while any {@code fleet.leaders}
-     * entry names a {@code tab}. A pane-placed member lands inside the focused tab rather than its
+     * or {@code fleet.collaborators} entry names a {@code tab}. A pane-placed member lands inside
-     * own, so it can land inside a lead's own labelled tab. {@link
+     * the focused tab rather than its own, so it can land inside a lead's or collaborator's own
-     * dev.ltms.fleet.herdr.LeadTabScanner} identifies a lead purely by that tab's label — it does
+     * labelled tab. {@link dev.ltms.fleet.herdr.LeadTabScanner} identifies a lead or collaborator
-     * not exclude the member space — so a member that ends up there would be read back as the lead
+     * purely by that tab's label — it does not exclude the member space — so a member that ends up
-     * and granted spawn/stop/send on the whole fleet.
+     * there would be read back as that lead or collaborator and granted that identity's authority.
      *
-     * <p>Only a leader with a non-blank {@code tab} is in scope: one with no {@code tab} feeds
+     * <p>Only an entry with a non-blank {@code tab} is in scope: one with no {@code tab} feeds
      * nothing into {@link dev.ltms.fleet.herdr.LeadTabScanner}, so it creates no hazard here.
      *
      * @throws IllegalStateException when any {@code profiles:} entry is pane-placed while any
-     *                               {@code fleet.leaders} entry names a non-blank {@code tab}
+     *                               {@code fleet.leaders} or {@code fleet.collaborators} entry
+     *                               names a non-blank {@code tab}
      */
     public void validatePanePlacementAgainstLeadTabs() {
-        if (fleet == null || fleet.leaders().isEmpty()) {
+        if (fleet == null) {
             return;
         }
         boolean anyLeaderHasTab = fleet.leaders().values().stream()
                 .anyMatch(leader -> leader != null && leader.tab() != null && !leader.tab().isBlank());
-        if (!anyLeaderHasTab) {
+        boolean anyCollaboratorHasTab = fleet.collaborators().values().stream()
+                .anyMatch(c -> c != null && c.tab() != null && !c.tab().isBlank());
+        if (!anyLeaderHasTab && !anyCollaboratorHasTab) {
             return;
         }
         List<String> bad = new ArrayList<>();
@@ -2771,10 +2907,11 @@ public record FleetConfig(
             return;
         }
         throw new IllegalStateException("refusing to start: profile(s) " + bad
-                + " use placement: pane while fleet.leaders names a tab. A pane-placed member can "
+                + " use placement: pane while fleet.leaders or fleet.collaborators names a tab. A "
-                + "land inside a lead's labelled tab and be read back as the lead, granted "
+                + "pane-placed member can land inside that labelled tab and be read back as the "
-                + "spawn/stop/send on the whole fleet. Set placement: tab for each named profile, "
+                + "lead or collaborator, granted that identity's authority. Set placement: tab for "
-                + "or remove the tab from every fleet.leaders entry.");
+                + "each named profile, or remove the tab from every fleet.leaders and "
+                + "fleet.collaborators entry.");
     }
 
     /**
@@ -2797,15 +2934,6 @@ public record FleetConfig(
         }
     }
 
-    /** Case-insensitive prefix test that tolerates a null/blank label. */
-    private static boolean startsWithIgnoreCase(String label, String prefix) {
-        if (label == null || prefix == null || prefix.isBlank()) {
-            return false;
-        }
-        String stripped = label.strip();
-        return stripped.regionMatches(true, 0, prefix, 0, prefix.length());
-    }
-
     /**
      * Reject a subscription profile whose {@code env:} block tries to reseat the Anthropic binding
      * (CB-542).
@@ -2890,8 +3018,14 @@ public record FleetConfig(
      * so duplicates are unrepresentable by construction once loaded — and {@link #load(Path)}
      * already rejects a duplicated slot name at parse time, before the map collapses.
      *
-     * @throws IllegalStateException when a slot names no profile or an unknown one, or when a lead
+     * <p>Also rejects a {@code fleet.collaborators} entry with no (or a blank) {@code tab}. A
-     *                               can be neither found nor created, naming the offending entry
+     * {@code profile}-less lead is still useful recognise-only — {@code tab} is the only field
+     * that matters to it either way. A collaborator carries no other field at all, so a blank
+     * {@code tab} leaves nothing for the entry to mean.
+     *
+     * @throws IllegalStateException when a slot names no profile or an unknown one, when a lead
+     *                               can be neither found nor created, or when a collaborator names
+     *                               no tab, naming the offending entry
      */
     public void validateMembers() {
         if (fleet == null) {
@@ -2929,6 +3063,16 @@ public record FleetConfig(
                         + "auto-launched, labelled) purely by its tab, so every entry must name one.");
             }
         });
+        fleet.collaborators().forEach((name, collaborator) -> {
+            if (collaborator == null) {
+                return;
+            }
+            if (collaborator.tab() == null || collaborator.tab().isBlank()) {
+                bad.add("fleet.collaborators." + name + " has no tab: — a collaborator is "
+                        + "recognised purely by its tab, and carries no other field, so every "
+                        + "entry must name one.");
+            }
+        });
         if (!bad.isEmpty()) {
             throw new IllegalStateException("refusing to start: " + String.join(" ", bad));
         }
@@ -2977,11 +3121,11 @@ public record FleetConfig(
      * Runs every validator this class declares — found by reflection, not by name.
      *
      * <p>fleetd ticket "central allow-list of usable models", follow-up: mutation testing found
-     * that although each of the six validators above was well pinned on its own, nothing proved
+     * that although each validator above was well pinned on its own, nothing proved
      * either real caller ({@code Fleetd.main} and {@link ConfigRef#reload()}) still
-     * invoked it — deleting a call site left the full suite green. The fix is not a seventh test
+     * invoked it — deleting a call site left the full suite green. The fix is not one more test
-     * per caller; a hand-maintained list of six names here would have the exact same defect its
+     * per caller; a hand-maintained list of names here would have the exact same defect its
-     * own javadoc would warn against: the seventh validator someone adds next month has no reason
+     * own javadoc would warn against: the next validator someone adds has no reason
      * to be added to it. So this method does not name any validator. It sweeps {@link
      * #getClass()}'s own public, no-argument, {@code void} methods whose name starts with {@code
      * "validate"} (excluding itself) and invokes every one it finds, via {@link
@@ -2990,7 +3134,7 @@ public record FleetConfig(
      * which it silently never runs.
      *
      * <p>{@code Fleetd.main} and {@link ConfigRef#reload()} each call this one method instead of
-     * the six individually — see the comments at those two call sites for why
+     * each validator individually — see the comments at those two call sites for why
      * each must run it.
      *
      * <p>Methods run in a fixed (alphabetical) order, so a config with more than one violation
@@ -3007,9 +3151,9 @@ public record FleetConfig(
     /**
      * The reflective sweep behind {@link #validateAll()}, kept as its own method — taking any
      * {@code target}, not just {@code this} — so a test can prove the MECHANISM is generic (it
-     * would sweep a seventh {@code validateXxx()} method added to any class, not just something
+     * would sweep any new {@code validateXxx()} method added to any class, not just something
-     * special-cased to today's six on {@link FleetConfig}) without needing to add a real, unwanted
+     * special-cased to the set {@link FleetConfig} declares today) without needing to add a real,
-     * seventh validator to this class just to exercise that claim. See {@code
+     * unwanted extra validator to this class just to exercise that claim. See {@code
      * FleetConfigValidateAllTest} for that proof.
      *
      * @param target an object whose public, no-argument, {@code void} methods named {@code

fleetd/src/main/java/dev/ltms/fleet/lead/LeadContextGauge.java

@@ -172,12 +172,6 @@ public final class LeadContextGauge {
      *                  {@code "claude"} (including {@code null}, meaning undetected) reports
      *                  {@link State#UNKNOWN} — this reader only understands Claude Code's own
      *                  transcript format
-     */
-    public Reading read(String configDir, String sessionId, String agentType) {
-        return read(configDir, sessionId, agentType, null);
-    }
-
-    /**
      * @param effectiveWindowTokens the caller's resolved effective auto-compact window for this
      *                              lead's own profile, or {@code null} when it cannot be resolved.
      *                              HIGH fires at {@link #HIGH_THRESHOLD_FRACTION} of this value;

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

@@ -689,8 +689,8 @@ public final class FleetMcp {
         if (!authorizationEnforced) {
             return null; // AuthorizationMode.UNENFORCED: authorization not enforced (fleetd #518)
         }
-        if (Authz.permits(caller, action, target)) {
+        if (Authz.permits(caller, action, target, Authz.NO_KNOWN_LEAD_OR_COLLABORATOR)) {
-            if (action != Authz.Action.READ) {
+            if (action != Authz.Action.READ && action != Authz.Action.TASK_READ) {
                 AuditLog.allowed(caller, action, target); // reads would drown the trail
             }
             return null;
@@ -1035,31 +1035,21 @@ public final class FleetMcp {
     }
 
     /**
-     * Which authorization action a {@code fleet_poll} call needs, decided by its arguments
+     * Which authorization action a {@code fleet_poll} call needs, decided by its arguments.
-     * (fleetd #272, widened by fleetd #421).
      *
-     * <p>{@code fleet_poll} is now <strong>three operations behind one tool name</strong>. With
+     * <p>{@code fleet_poll} is <strong>three operations behind one tool name</strong>. With
      * {@code ticket} it observes an async delegation and changes nothing, which is a {@link
-     * Authz.Action#READ}. With {@code target} it calls {@link MessageService#drainReplies} on that
+     * Authz.Action#TASK_READ}. With {@code target} it calls {@link MessageService#drainReplies} on
-     * session -- the replies are removed from the inbox and a second call returns nothing -- so it
+     * that session -- the replies are removed from the inbox and a second call returns nothing --
-     * is a {@link Authz.Action#DRAIN}, the same gate {@code fleet_ack} already uses for removing a
+     * so it is a {@link Authz.Action#DRAIN}, the same gate {@code fleet_ack} already uses for
-     * single message, and the same one the REST path uses at {@code FleetApp.drainReplies}. With
+     * removing a single message, and the same one the REST path uses at
-     * {@code coordId} it reads (never acks) this daemon's own held lead-to-lead mail, which is a
+     * {@code FleetApp.drainReplies}. With {@code coordId} it reads (never acks) this daemon's own
-     * {@link Authz.Action#COORD_READ} -- <strong>not</strong> {@code READ}, even though nothing is
+     * held lead-to-lead mail, which is a {@link Authz.Action#COORD_READ} -- <strong>not</strong>
-     * consumed: {@code READ}'s grant is open to every authenticated role on the premise that the
+     * {@code TASK_READ} or {@code READ}: a lead-to-lead body is a different inbox from either, and
-     * roster carries no secrets, and a lead-to-lead body is not the roster. Mapping a non-destructive
+     * folding it into either would let any worker or architect read every peer lead's mail in full.
-     * peer-mail read to {@code READ} would let any worker read every peer lead's mail in full.
-     *
-     * <p>Before this method existed (fleetd #272) the handler passed a constant {@code READ} for
-     * both of the original branches. {@code READ} is open to every authenticated role, so any
-     * worker could read a peer's id out of {@code fleet_list} and destroy the replies that peer had
-     * queued for the primary. The gate failed open, and it did so because the required action is a
-     * function of the arguments while the handler chose it before looking at them.
      *
      * <p>The choice lives in this method, and not inline in the handler, so that a test can assert
-     * the mapping the handler actually uses. {@code FleetMcpAuthzTest} already checked every
+     * the mapping the handler actually uses.
-     * {@link Authz.Action} against every {@link Role} and passed throughout -- it tested the policy
-     * table, which was correct, while the defect was in which action the caller handed it.
      *
      * <p>Checked first, and exclusively of {@code target}: a call naming {@code coordId} is reading
      * a different inbox entirely (this daemon's own lead channel, never a worker's), so it takes
@@ -1072,7 +1062,30 @@ public final class FleetMcp {
         if (!isBlank(coordId)) {
             return Authz.Action.COORD_READ;
         }
-        return isBlank(target) ? Authz.Action.READ : Authz.Action.DRAIN;
+        return isBlank(target) ? Authz.Action.TASK_READ : Authz.Action.DRAIN;
+    }
+
+    /**
+     * Which authorization action a {@code fleet_send} call needs, decided by its arguments.
+     *
+     * <p>{@code fleet_send} is three call shapes behind one tool name, mirroring {@link
+     * #pollAction}. With {@code coordId} it addresses a peer lead on another daemon over the
+     * coordination broker, which is {@link Authz.Action#COORD_SEND}. With {@code turnId} it
+     * resolves a worker's blocked {@code fleet_ask} and resumes that turn, which is {@link
+     * Authz.Action#ANSWER}. Otherwise it delivers to a local session by {@code sessionId}, which is
+     * the plain {@link Authz.Action#SEND}.
+     *
+     * <p>Checked in the same order the handler branches: {@code coordId} first and exclusively of
+     * {@code turnId}, matching {@link #sendToLead}'s own mutual-exclusion check.
+     *
+     * @param coordId the {@code coordId} argument of the call, or {@code null}/blank when absent
+     * @param turnId  the {@code turnId} argument of the call, or {@code null}/blank when absent
+     */
+    static Authz.Action sendAction(String coordId, String turnId) {
+        if (!isBlank(coordId)) {
+            return Authz.Action.COORD_SEND;
+        }
+        return isBlank(turnId) ? Authz.Action.SEND : Authz.Action.ANSWER;
     }
 
     /**
@@ -1097,10 +1110,11 @@ public final class FleetMcp {
      */
     private static Authz.Action authzAction(FleetTool tool, Map<String, Object> arguments) {
         return switch (tool) {
-            case SEND -> Authz.Action.SEND;
+            case SEND -> sendAction(str(arguments, "coordId"), str(arguments, "turnId"));
             case REPLY -> Authz.Action.REPLY;
             case ASK -> Authz.Action.ASK;
-            case STATUS, LIST, PROFILES, WHOAMI -> Authz.Action.READ;
+            case STATUS -> Authz.Action.TASK_READ;
+            case LIST, PROFILES, WHOAMI -> Authz.Action.READ;
             case POLL -> pollAction(str(arguments, "target"), str(arguments, "coordId"));
             case ACK -> Authz.Action.DRAIN;
             case SPAWN -> Authz.Action.SPAWN;
@@ -1295,6 +1309,18 @@ public final class FleetMcp {
             }
             return text(json(m));
         }
+        if (caller.isCollaborator()) {
+            // A collaborator's name is its slot in the collaborators: registry; sessionId is its
+            // pane so a peer knows where to reach it. No leader key: a collaborator is not a
+            // primary for authorization, unlike a lead.
+            if (caller.name() != null) {
+                m.put("collaborator", caller.name());
+            }
+            if (caller.terminal() != null) {
+                m.put("sessionId", caller.terminal());
+            }
+            return text(json(m));
+        }
         if (!caller.isWorker()) {
             // CB-530: which lead, once more than one pane is configured as one. `role` deliberately
             // still reads "primary" — the fallback ladder in CLAUDE.md keys on it, and a lead IS a

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

@@ -49,22 +49,52 @@ import java.util.stream.Collectors;
  */
 public final class FleetApp {
 
-    /** The authorization action the matching route handler hands to {@link #allow}. */
+    /**
+     * The authorization action the matching route handler hands to {@link #allow}, for a route
+     * whose action does not depend on the request body.
+     */
     static Authz.Action routeAction(String route) {
+        return routeAction(route, null);
+    }
+
+    /**
+     * As above, plus the one route whose action depends on the body: {@code POST
+     * /sessions/{id}/message} carries a {@code turnId} (the answer-a-blocked-worker shape) or not
+     * (a plain delivery), mirroring {@code FleetMcp#sendAction}'s split of the same two call
+     * shapes over MCP. {@code turnId} is ignored by every other route.
+     *
+     * @param turnId the request body's {@code turnId}, or {@code null}/blank when absent or not
+     *               applicable to this route
+     */
+    static Authz.Action routeAction(String route, String turnId) {
         return switch (route) {
             case "GET /metrics" -> Authz.Action.METRICS;
             case "POST /members" -> Authz.Action.SPAWN;
             case "DELETE /members/{paneId}" -> Authz.Action.STOP;
-            case "POST /sessions/{id}/message" -> Authz.Action.SEND;
+            case "POST /sessions/{id}/message" -> turnId == null || turnId.isBlank()
+                    ? Authz.Action.SEND : Authz.Action.ANSWER;
             case "POST /sessions/{id}/reply" -> Authz.Action.REPLY;
             case "GET /sessions/{id}/replies" -> Authz.Action.DRAIN;
             case "POST /sessions/{id}/ask" -> Authz.Action.ASK;
             case "GET /sessions", "GET /agents", "GET /members", "GET /profiles",
-                    "GET /member-credentials", "GET /sessions/{id}/status", "GET /tasks/{ticket}" -> Authz.Action.READ;
+                    "GET /member-credentials" -> Authz.Action.READ;
+            case "GET /sessions/{id}/status", "GET /tasks/{ticket}" -> Authz.Action.TASK_READ;
             default -> throw new IllegalArgumentException("route has no authorization gate: " + route);
         };
     }
 
+    /**
+     * The second gate for {@code POST /sessions/{id}/message}: checked only when {@code turnId}
+     * is present and non-blank, against {@link Authz.Action#ANSWER}. A request with no {@code
+     * turnId} passes this gate unconditionally, without consulting {@code permit} at all, having
+     * already cleared the coarse {@link Authz.Action#SEND} grant checked ahead of it.
+     *
+     * @param permit reports whether the caller holds the named grant
+     */
+    static boolean answerGatePasses(String turnId, Predicate<Authz.Action> permit) {
+        return turnId == null || turnId.isBlank() || permit.test(Authz.Action.ANSWER);
+    }
+
     /** Default blocking window for a message; kept under typical HTTP idle timeouts. */
     private static final long DEFAULT_MESSAGE_TIMEOUT_MS = 25_000;
     private static final long MAX_MESSAGE_TIMEOUT_MS = 120_000;
@@ -236,6 +266,17 @@ public final class FleetApp {
         return app;
     }
 
+    /**
+     * The authorization decision behind {@link #allow}, taking the caller directly rather than
+     * pulling it from a servlet {@link Context} — unit-testable without fabricating a live
+     * request, the same reason {@code FleetMcp#denyFor} is split from {@code FleetMcp#deny}. The
+     * classifier is the same shared instance {@link #allow} would use, so a test calling this
+     * exercises the real production gate, not a test-supplied stand-in.
+     */
+    static boolean permitsFor(Principal caller, Authz.Action action, String target) {
+        return Authz.permits(caller, action, target, Authz.NO_KNOWN_LEAD_OR_COLLABORATOR);
+    }
+
     /**
      * Gate a handler on the CB-505 authorization table. Returns {@code true} when the request may
      * proceed; otherwise writes the error response and returns {@code false}.
@@ -249,8 +290,9 @@ public final class FleetApp {
             return true; // legacy: authorization not enforced
         }
         Principal caller = ctx.attribute(CALLER);
-        if (Authz.permits(caller, action, target)) {
+        if (permitsFor(caller, action, target)) {
-            if (action != Authz.Action.READ && action != Authz.Action.METRICS) {
+            if (action != Authz.Action.READ && action != Authz.Action.METRICS
+                    && action != Authz.Action.TASK_READ) {
                 AuditLog.allowed(caller, action, target); // reads would drown the trail
             }
             return true;
@@ -603,26 +645,37 @@ public final class FleetApp {
      * status-gated injector and block until the worker returns a structured {@code fleet_reply}.
      * Times out with a typed 202 (working / queued / busy) rather than an error — the message may
      * still land.
+     *
+     * <p>Two call shapes share this route, exactly as {@code fleet_send} does over MCP (see
+     * {@code FleetMcp#sendAction}): a plain delivery to {@code id}, and -- when the body carries
+     * {@code turnId} -- resolving a worker's blocked question. The coarse {@link
+     * Authz.Action#SEND} grant is checked first, before the body is read at all; only once that
+     * passes is the body parsed, and a present {@code turnId} is then checked again against
+     * {@link Authz.Action#ANSWER}. A body that fails to parse is rejected with 400 and reaches
+     * neither {@code messages.answer} nor {@code messages.send}.
      */
     private void sendMessage(Context ctx) {
         String id = ctx.pathParam("id");
         if (!allow(ctx, routeAction("POST /sessions/{id}/message"), id)) {
             return;
         }
-        String content;
+        JsonNode body;
-        String turnId;
-        long timeout;
-        boolean wait;
         try {
-            JsonNode body = mapper.readTree(ctx.body());
+            body = mapper.readTree(ctx.body());
-            content = body.path("content").asText("");
-            turnId = body.path("turnId").asText(null);
-            timeout = body.path("timeoutMs").asLong(DEFAULT_MESSAGE_TIMEOUT_MS);
-            wait = body.path("wait").asBoolean(true); // default: block for the reply (CB-104)
         } catch (Exception e) {
+            body = null;
+        }
+        if (body == null) {
             ctx.status(400).json(Map.of("error", "bad_request", "detail", "body must be JSON"));
             return;
         }
+        String turnId = body.path("turnId").asText(null);
+        if (!answerGatePasses(turnId, action -> allow(ctx, action, id))) {
+            return;
+        }
+        String content = body.path("content").asText("");
+        long timeout = body.path("timeoutMs").asLong(DEFAULT_MESSAGE_TIMEOUT_MS);
+        boolean wait = body.path("wait").asBoolean(true); // default: block for the reply (CB-104)
         if (content.isBlank()) {
             ctx.status(400).json(Map.of("error", "bad_request", "detail", "content is required"));
             return;

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

@@ -0,0 +1,166 @@
+package dev.ltms.fleet;
+
+import dev.ltms.fleet.auth.Authz;
+import dev.ltms.fleet.auth.Principal;
+import dev.ltms.fleet.config.ConfigRef;
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.guard.SubscriptionGuard;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
+import dev.ltms.fleet.mcp.FleetMcp;
+import dev.ltms.fleet.msg.ReplyInbox;
+import io.javalin.Javalin;
+import io.modelcontextprotocol.spec.McpSchema;
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+
+import java.lang.reflect.Method;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.Executors;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.function.LongSupplier;
+
+import static org.junit.jupiter.api.Assertions.assertNotNull;
+import static org.junit.jupiter.api.Assertions.assertNull;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * Asserts that the {@link FleetMcp} built by {@link FleetdAssembly#assembleAndStart} applies the
+ * authorization table: a worker is refused {@code SPAWN}, and the primary is allowed it.
+ */
+class FleetdAssemblyAuthorizationModeTest {
+
+    private static final class TestResourcePorts implements ResourcePorts {
+        final FakeHerdr herdr = new FakeHerdr();
+        Runnable shutdownHook;
+
+        @Override
+        public Map<String, String> environment() {
+            return Map.of();
+        }
+
+        @Override
+        public HerdrClient connectHerdr(Path socketPath) {
+            return herdr;
+        }
+
+        @Override
+        public Fleetd.AmqpOpener replyInboxOpener() {
+            return (uri, prefetch) -> new ReplyInbox() {
+                @Override public void own(String target) { }
+                @Override public void release(String target) { }
+                @Override public void publish(String target, String msgId, String content) { }
+                @Override public List<InboxMessage> peek(String target) { return List.of(); }
+                @Override public boolean ack(String target, String msgId) { return false; }
+            };
+        }
+
+        @Override
+        public Fleetd.LeadMailboxOpener leadMailboxOpener() {
+            return (uri, selfCoordId, prefetch) -> {
+                throw new UnsupportedOperationException(
+                        "leadMailboxOpener must not be called — no coordinator: block is configured");
+            };
+        }
+
+        @Override
+        public LongSupplier nanoClock() {
+            return System::nanoTime;
+        }
+
+        @Override
+        public LongSupplier wallClockNanos() {
+            return System::nanoTime;
+        }
+
+        @Override
+        public ScheduledExecutorService newScheduler(String purpose) {
+            return Executors.newSingleThreadScheduledExecutor();
+        }
+
+        @Override
+        public void addShutdownHook(Runnable hook) {
+            shutdownHook = hook;
+        }
+
+        @Override
+        public void startHttp(Javalin app, String host, int port) {
+            // Binding a real port would clash with any daemon already listening on it.
+        }
+
+        @Override
+        public Runnable herdrPollWait() {
+            return () -> {
+                throw new UnsupportedOperationException("FakeHerdr is healthy; no poll wait is expected");
+            };
+        }
+    }
+
+    private TestResourcePorts ports;
+
+    @AfterEach
+    void tearDown() {
+        if (ports != null && ports.shutdownHook != null) {
+            ports.shutdownHook.run();
+        }
+    }
+
+    private static FleetConfig writeConfig(Path dir) throws Exception {
+        Path file = dir.resolve("fleetd.yaml");
+        Files.writeString(file, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                idleSleepGuard:
+                  enabled: false
+                health:
+                  enabled: false
+                broker:
+                  uri: "amqp://fake-test-broker/vh"
+                """);
+        return FleetConfig.load(file);
+    }
+
+    private FleetMcp assemble(Path dir) throws Exception {
+        FleetConfig cfg = writeConfig(dir);
+        ports = new TestResourcePorts();
+        FleetdRuntime runtime = FleetdAssembly.assembleAndStart(new AssemblyInputs(cfg,
+                new ConfigRef(dir.resolve("fleetd.yaml"), cfg), new SubscriptionGuard(cfg.guard().hostSet())), ports);
+        return runtime.mcp();
+    }
+
+    /**
+     * Invokes {@code FleetMcp#denyFor}, which is package-private to {@code dev.ltms.fleet.mcp}
+     * while this test is in {@code dev.ltms.fleet}. Nothing here catches a missing method: if
+     * {@code denyFor} is renamed or removed, {@link NoSuchMethodException} propagates and the
+     * test fails.
+     */
+    private static McpSchema.CallToolResult denyFor(FleetMcp mcp, Principal caller, Authz.Action action,
+                                                      String target) throws Exception {
+        Method m = FleetMcp.class.getDeclaredMethod("denyFor", Principal.class, Authz.Action.class, String.class);
+        m.setAccessible(true);
+        return (McpSchema.CallToolResult) m.invoke(mcp, caller, action, target);
+    }
+
+    @Test
+    void productionBootPathRefusesAnUnauthorizedCallerThroughTheAssembledFleetMcp(@TempDir Path dir)
+            throws Exception {
+        FleetMcp mcp = assemble(dir);
+
+        McpSchema.CallToolResult deniedForWorker = denyFor(mcp, Principal.worker("term_a", 200),
+                Authz.Action.SPAWN, "term_a");
+        assertNotNull(deniedForWorker,
+                "a worker must not be able to fleet_spawn through the assembled FleetMcp");
+        assertTrue(deniedForWorker.isError(), "a refusal is returned as an MCP tool error");
+
+        McpSchema.CallToolResult allowedForPrimary = denyFor(mcp, Principal.primary(100),
+                Authz.Action.SPAWN, "term_a");
+        assertNull(allowedForPrimary,
+                "control: the primary must still be allowed to fleet_spawn — otherwise the worker "
+                        + "refusal above would pass even with the gate wired backwards");
+    }
+}

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

@@ -0,0 +1,201 @@
+package dev.ltms.fleet;
+
+import dev.ltms.fleet.config.ConfigRef;
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.guard.SubscriptionGuard;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
+import dev.ltms.fleet.herdr.LeadTabScanner;
+import dev.ltms.fleet.msg.LeadChannelHandle;
+import dev.ltms.fleet.msg.LeadCoordLoop;
+import dev.ltms.fleet.msg.LeadMessage;
+import dev.ltms.fleet.msg.ReplyInbox;
+import io.javalin.Javalin;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+
+import java.lang.reflect.Field;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.concurrent.Executors;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.function.LongSupplier;
+import java.util.function.Supplier;
+
+import static org.junit.jupiter.api.Assertions.assertInstanceOf;
+import static org.junit.jupiter.api.Assertions.assertNotNull;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #670 — pins the {@code excludedWorkspaceLabels} argument {@link FleetdAssembly}'s
+ * production boot path passes to {@link LeadTabScanner} at {@code FleetdAssembly.java:265}
+ * ({@code Set.of()}).
+ *
+ * <p>{@code LeadTabScannerTest} already covers this constructor parameter, but it builds its own
+ * {@link LeadTabScanner} with its own set, so it tests the seam and proves nothing about the
+ * producer. This test instead reaches the exact object {@link FleetdAssembly#assembleAndStart}
+ * builds: a {@code fleet.leaders:} block makes the assembly construct a real
+ * {@link LeadTabScanner} for its local {@code leads} supplier, and a {@code coordinator:} block
+ * makes it hand that same supplier instance to {@link LeadCoordLoop} (fleetd #637), which stores
+ * it as a field. Reflection recovers it from there, and then from the scanner itself, so the
+ * assertion is against the real production argument rather than a copy built for this test.
+ */
+class FleetdAssemblyLeadTabScannerExclusionTest {
+
+    private static final class FakeLeadChannel implements LeadChannelHandle {
+        @Override
+        public void publish(String toCoordId, LeadMessage message) {
+        }
+
+        @Override
+        public List<LeadMessage> peek() {
+            return List.of();
+        }
+
+        @Override
+        public void ack(String msgId) {
+        }
+
+        @Override
+        public String selfCoordId() {
+            return "test-lead";
+        }
+
+        @Override
+        public boolean heldDurable() {
+            return true;
+        }
+
+        @Override
+        public MailboxState inspect(String coordId) {
+            return MailboxState.unknown(coordId);
+        }
+
+        @Override
+        public void close() {
+        }
+    }
+
+    private static final class TestResourcePorts implements ResourcePorts {
+        final FakeHerdr herdr = new FakeHerdr();
+        Runnable shutdownHook;
+
+        @Override
+        public Map<String, String> environment() {
+            return Map.of();
+        }
+
+        @Override
+        public HerdrClient connectHerdr(Path socketPath) {
+            return herdr;
+        }
+
+        @Override
+        public Fleetd.AmqpOpener replyInboxOpener() {
+            return (uri, prefetch) -> new ReplyInbox() {
+                @Override public void own(String target) { }
+                @Override public void release(String target) { }
+                @Override public void publish(String target, String msgId, String content) { }
+                @Override public List<InboxMessage> peek(String target) { return List.of(); }
+                @Override public boolean ack(String target, String msgId) { return false; }
+            };
+        }
+
+        @Override
+        public Fleetd.LeadMailboxOpener leadMailboxOpener() {
+            return (uri, selfCoordId, prefetch) -> new FakeLeadChannel();
+        }
+
+        @Override
+        public LongSupplier nanoClock() {
+            return System::nanoTime;
+        }
+
+        @Override
+        public LongSupplier wallClockNanos() {
+            return System::nanoTime;
+        }
+
+        @Override
+        public ScheduledExecutorService newScheduler(String purpose) {
+            return Executors.newSingleThreadScheduledExecutor();
+        }
+
+        @Override
+        public void addShutdownHook(Runnable hook) {
+            shutdownHook = hook;
+        }
+
+        @Override
+        public void startHttp(Javalin app, String host, int port) {
+            // Do not bind a real port in this assembly test.
+        }
+
+        @Override
+        public Runnable herdrPollWait() {
+            return () -> {
+                throw new UnsupportedOperationException("FakeHerdr is healthy; no poll wait is expected");
+            };
+        }
+    }
+
+    private static FleetConfig writeConfig(Path dir) throws Exception {
+        Path file = dir.resolve("fleetd.yaml");
+        Files.writeString(file, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                idleSleepGuard:
+                  enabled: false
+                coordinator:
+                  uri: "amqp://fake-lead-broker/vh"
+                  selfId: "test-lead"
+                fleet:
+                  leaders:
+                    primary:
+                      tab: "lead: primary"
+                      profile: sonnet
+                profiles:
+                  sonnet:
+                    subscription: true
+                    argv: ["ccs", "sonnet"]
+                """);
+        return FleetConfig.load(file);
+    }
+
+    @Test
+    void productionBootPathPassesNoExcludedWorkspaceLabels(@TempDir Path dir) throws Exception {
+        FleetConfig cfg = writeConfig(dir);
+        TestResourcePorts ports = new TestResourcePorts();
+        FleetdRuntime runtime = FleetdAssembly.assembleAndStart(new AssemblyInputs(cfg,
+                new ConfigRef(dir.resolve("fleetd.yaml"), cfg), new SubscriptionGuard(cfg.guard().hostSet())), ports);
+        try {
+            LeadCoordLoop coordLoop = runtime.leadCoordLoop();
+            assertNotNull(coordLoop, "control: a configured coordinator: block must build LeadCoordLoop");
+
+            Field leadsField = LeadCoordLoop.class.getDeclaredField("leads");
+            leadsField.setAccessible(true);
+            @SuppressWarnings("unchecked")
+            Supplier<Map<String, String>> leads = (Supplier<Map<String, String>>) leadsField.get(coordLoop);
+
+            assertInstanceOf(LeadTabScanner.class, leads,
+                    "control: a non-empty fleet.leaders: block must make FleetdAssembly build a real "
+                            + "LeadTabScanner for its `leads` supplier, not the Map::of fallback — "
+                            + "otherwise this test would pass for the wrong reason");
+
+            Field excludedField = LeadTabScanner.class.getDeclaredField("excludedWorkspaceLabels");
+            excludedField.setAccessible(true);
+            Set<?> excluded = (Set<?>) excludedField.get(leads);
+
+            assertTrue(excluded.isEmpty(),
+                    "FleetdAssembly.java:265 must pass an empty excludedWorkspaceLabels to "
+                            + "LeadTabScanner — scanning member tabs would demote the lead to a worker");
+        } finally {
+            assertNotNull(ports.shutdownHook, "control: assembly must capture its shutdown hook");
+            ports.shutdownHook.run();
+        }
+    }
+}

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

@@ -0,0 +1,190 @@
+package dev.ltms.fleet;
+
+import dev.ltms.fleet.config.ConfigRef;
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.guard.SubscriptionGuard;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
+import dev.ltms.fleet.inject.Injector;
+import dev.ltms.fleet.inject.StatusPoller;
+import dev.ltms.fleet.msg.LeadChannelHandle;
+import dev.ltms.fleet.msg.LeadCoordLoop;
+import dev.ltms.fleet.msg.LeadMessage;
+import dev.ltms.fleet.msg.ReplyInbox;
+import dev.ltms.fleet.msg.ReplyPushLoop;
+import io.javalin.Javalin;
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+
+import java.lang.reflect.Field;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.List;
+import java.util.Map;
+import java.util.concurrent.Executors;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.function.LongSupplier;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertNotNull;
+
+/**
+ * Asserts that the assembled loops use the production reminder, coordination, and delivery timing
+ * defaults when no {@code primary:} block configures the reply-push values.
+ */
+class FleetdAssemblyTimingDefaultsTest {
+
+    private static final class FakeLeadChannel implements LeadChannelHandle {
+        @Override
+        public void publish(String toCoordId, LeadMessage message) {
+        }
+
+        @Override
+        public List<LeadMessage> peek() {
+            return List.of();
+        }
+
+        @Override
+        public void ack(String msgId) {
+        }
+
+        @Override
+        public String selfCoordId() {
+            return "test-lead";
+        }
+
+        @Override
+        public boolean heldDurable() {
+            return true;
+        }
+
+        @Override
+        public MailboxState inspect(String coordId) {
+            return MailboxState.unknown(coordId);
+        }
+
+        @Override
+        public void close() {
+        }
+    }
+
+    private static final class TestResourcePorts implements ResourcePorts {
+        final FakeHerdr herdr = new FakeHerdr();
+        Runnable shutdownHook;
+
+        @Override
+        public Map<String, String> environment() {
+            return Map.of();
+        }
+
+        @Override
+        public HerdrClient connectHerdr(Path socketPath) {
+            return herdr;
+        }
+
+        @Override
+        public Fleetd.AmqpOpener replyInboxOpener() {
+            return (uri, prefetch) -> new ReplyInbox() {
+                @Override public void own(String target) { }
+                @Override public void release(String target) { }
+                @Override public void publish(String target, String msgId, String content) { }
+                @Override public List<InboxMessage> peek(String target) { return List.of(); }
+                @Override public boolean ack(String target, String msgId) { return false; }
+            };
+        }
+
+        @Override
+        public Fleetd.LeadMailboxOpener leadMailboxOpener() {
+            return (uri, selfCoordId, prefetch) -> new FakeLeadChannel();
+        }
+
+        @Override
+        public LongSupplier nanoClock() {
+            return System::nanoTime;
+        }
+
+        @Override
+        public LongSupplier wallClockNanos() {
+            return System::nanoTime;
+        }
+
+        @Override
+        public ScheduledExecutorService newScheduler(String purpose) {
+            return Executors.newSingleThreadScheduledExecutor();
+        }
+
+        @Override
+        public void addShutdownHook(Runnable hook) {
+            shutdownHook = hook;
+        }
+
+        @Override
+        public void startHttp(Javalin app, String host, int port) {
+        }
+
+        @Override
+        public Runnable herdrPollWait() {
+            return () -> {
+                throw new UnsupportedOperationException("FakeHerdr is healthy; no poll wait is expected");
+            };
+        }
+    }
+
+    private TestResourcePorts ports;
+
+    @AfterEach
+    void tearDown() {
+        if (ports != null && ports.shutdownHook != null) {
+            ports.shutdownHook.run();
+        }
+    }
+
+    private static FleetConfig writeConfig(Path dir) throws Exception {
+        Path file = dir.resolve("fleetd.yaml");
+        Files.writeString(file, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                idleSleepGuard:
+                  enabled: false
+                coordinator:
+                  uri: "amqp://fake-lead-broker/vh"
+                  selfId: "test-lead"
+                """);
+        return FleetConfig.load(file);
+    }
+
+    private FleetdRuntime assemble(Path dir) throws Exception {
+        FleetConfig cfg = writeConfig(dir);
+        ports = new TestResourcePorts();
+        return FleetdAssembly.assembleAndStart(new AssemblyInputs(cfg,
+                new ConfigRef(dir.resolve("fleetd.yaml"), cfg), new SubscriptionGuard(cfg.guard().hostSet())), ports);
+    }
+
+    private static long longField(Object target, String name) throws Exception {
+        Field field = target.getClass().getDeclaredField(name);
+        field.setAccessible(true);
+        return field.getLong(target);
+    }
+
+    @Test
+    void productionBootPathUsesTheExpectedLoopTimingDefaults(@TempDir Path dir) throws Exception {
+        FleetdRuntime runtime = assemble(dir);
+
+        ReplyPushLoop pushLoop = runtime.pushLoop();
+        assertEquals(5, longField(pushLoop, "maxReminders"),
+                "without primary:, ReplyPushLoop must stop after five reminder attempts");
+        assertEquals(15_000L, longField(pushLoop, "backoffMs"),
+                "without primary:, ReplyPushLoop must wait fifteen seconds before the next reminder");
+
+        LeadCoordLoop leadCoordLoop = runtime.leadCoordLoop();
+        assertNotNull(leadCoordLoop, "control: coordinator: must build LeadCoordLoop");
+        assertEquals(3_000L, longField(leadCoordLoop, "intervalMs"),
+                "LeadCoordLoop must poll for peer-lead mail every three seconds");
+
+        StatusPoller poller = runtime.poller();
+        assertEquals(Injector.POLL_INTERVAL_MILLIS, longField(poller, "intervalMillis"),
+                "StatusPoller must use Injector's delivery poll interval");
+    }
+}

fleetd/src/test/java/dev/ltms/fleet/auth/AuthzTest.java

@@ -14,6 +14,7 @@ class AuthzTest {
     private static final Principal ANON = Principal.anonymous();
     private static final Principal ARCH_DESIGN = Principal.architect("lead-designer", "term_design", 400);
     private static final Principal ARCH_OTHER = Principal.architect("reviewer", "term_review", 500);
+    private static final Principal COLLABORATOR = Principal.collaborator("ops", "term_collab", 600);
 
     @Test
     void anonymousIsAuthorizedForNothing() {
@@ -38,6 +39,37 @@ class AuthzTest {
         }
     }
 
+    /**
+     * {@code fleet_send} is three call shapes behind one action name until {@code
+     * FleetMcp#sendAction} picks one: a plain local {@link Authz.Action#SEND}, the {@code coordId}
+     * route ({@link Authz.Action#COORD_SEND}), and the {@code turnId} answer form ({@link
+     * Authz.Action#ANSWER}). All three carry the same grant as the undivided action did — a worker
+     * is excluded from every one, exactly as it was excluded from the one combined action before.
+     */
+    @Test
+    void theThreeSendShapesCarryTheSameGrantAsTheOldUndividedAction() {
+        for (Authz.Action a : new Authz.Action[]{SEND, COORD_SEND, ANSWER}) {
+            assertTrue(Authz.permits(PRIMARY, a, "term_a"), "the primary may " + a);
+            assertTrue(Authz.permits(ARCH_DESIGN, a, "term_a"), "an architect may " + a);
+            assertFalse(Authz.permits(WORKER_A, a, "term_a"),
+                    "a worker performing " + a + " would be escalating into the orchestrator role");
+            assertFalse(Authz.permits(ANON, a, "term_a"));
+        }
+    }
+
+    /**
+     * {@code fleet_poll{ticket}} and {@code fleet_status} are {@link Authz.Action#TASK_READ}, split
+     * out of the roster-only {@link Authz.Action#READ} (fleetd #678). The grant is unchanged from
+     * what the undivided {@code READ} action gave every one of these callers.
+     */
+    @Test
+    void taskReadCarriesTheSameGrantReadDidBeforeTheSplit() {
+        assertTrue(Authz.permits(PRIMARY, TASK_READ, null));
+        assertTrue(Authz.permits(WORKER_A, TASK_READ, null));
+        assertTrue(Authz.permits(ARCH_DESIGN, TASK_READ, null));
+        assertFalse(Authz.permits(ANON, TASK_READ, null));
+    }
+
     @Test
     void aWorkerMayReplyAndAskOnlyAsItself() {
         assertTrue(Authz.permits(WORKER_A, REPLY, "term_a"));
@@ -135,4 +167,87 @@ class AuthzTest {
         assertFalse(Authz.isUnauthenticated(WORKER_A));
         assertFalse(Authz.isUnauthenticated(PRIMARY));
     }
+
+    // ── the collaborator matrix ─────────────────────────────────────────────────────────────────
+
+    /**
+     * {@code SEND} for a collaborator is the one grant that is conditional rather than fixed:
+     * flipping only the classifier's answer for the target flips only this outcome.
+     */
+    @Test
+    void aCollaboratorMaySendOnlyWhenTheClassifierAcceptsTheTarget() {
+        assertTrue(Authz.permits(COLLABORATOR, SEND, "term_lead", target -> true),
+                "the classifier accepting the target must grant SEND");
+        assertFalse(Authz.permits(COLLABORATOR, SEND, "term_lead", target -> false),
+                "the classifier refusing the target must deny SEND");
+        assertFalse(Authz.permits(COLLABORATOR, SEND, "term_lead"),
+                "the real production classifier recognises no terminal yet, so SEND is refused today");
+    }
+
+    /**
+     * Control for the test above: every other action's result for a collaborator does not move
+     * when the classifier does. Only {@code SEND} is wired to it.
+     */
+    @Test
+    void theClassifierMovesOnlySendForACollaborator() {
+        for (Authz.Action a : Authz.Action.values()) {
+            if (a == SEND) {
+                continue;
+            }
+            assertEquals(
+                    Authz.permits(COLLABORATOR, a, "term_lead"),
+                    Authz.permits(COLLABORATOR, a, "term_lead", target -> true),
+                    a + " must not depend on the classifier at all");
+        }
+    }
+
+    @Test
+    void aCollaboratorMayReadAndScrapeMetrics() {
+        assertTrue(Authz.permits(COLLABORATOR, READ, null));
+        assertTrue(Authz.permits(COLLABORATOR, METRICS, null));
+    }
+
+    @Test
+    void aCollaboratorMayReplyAndAskOnlyAsItsOwnPane() {
+        assertTrue(Authz.permits(COLLABORATOR, REPLY, "term_collab"),
+                "its own pane is its own");
+        assertTrue(Authz.permits(COLLABORATOR, ASK, "term_collab"));
+
+        assertFalse(Authz.permits(COLLABORATOR, REPLY, "term_design"),
+                "a collaborator must not reply on another pane");
+        assertFalse(Authz.permits(COLLABORATOR, REPLY, null),
+                "an absent target must not pass the own-session rule");
+    }
+
+    /**
+     * Every action denied to a collaborator, asserted denied even when the classifier would
+     * accept any target — proving none of these is actually gated on the classifier at all.
+     */
+    @Test
+    void aCollaboratorIsDeniedLifecycleCoordinationAndTicketPolling() {
+        for (Authz.Action a : new Authz.Action[]{SPAWN, STOP, DRAIN, HANDOVER, ANSWER, COORD_SEND,
+                COORD_READ, TASK_READ}) {
+            assertFalse(Authz.permits(COLLABORATOR, a, "term_lead", target -> true),
+                    "a collaborator must not " + a + " even when the classifier accepts every target");
+        }
+    }
+
+    @Test
+    void aCollaboratorIsNotCountedAsPrimaryWorkerOrArchitect() {
+        assertFalse(COLLABORATOR.isPrimary());
+        assertFalse(COLLABORATOR.isWorker());
+        assertFalse(COLLABORATOR.isArchitect());
+        assertTrue(COLLABORATOR.isCollaborator());
+    }
+
+    /**
+     * A collaborator is never spawned, so it must not be enrolled in the presence map as an
+     * available member. Control: both a worker and an architect — which ARE spawned — still are.
+     */
+    @Test
+    void isSpawnedMemberIsFalseForACollaboratorButTrueForAWorkerAndAnArchitect() {
+        assertFalse(COLLABORATOR.isSpawnedMember());
+        assertTrue(WORKER_A.isSpawnedMember());
+        assertTrue(ARCH_DESIGN.isSpawnedMember());
+    }
 }

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

@@ -676,12 +676,8 @@ class FleetConfigTest {
         assertEquals(5, hb.quietNudgeCap());
     }
 
-    /**
-     * The hazard the guard exists for: fleetd writes worker tab labels and reads lead tab labels.
-     * Overlap the two and every worker it spawns is read back as a lead.
-     */
     @Test
-    void aLeadPrefixThatAProfileTabLabelOverrideAlsoMatchesRefusesToStart(@TempDir Path dir)
+    void aProfileTabLabelOverrideMatchingALeadTabRefusesToStart(@TempDir Path dir)
             throws Exception {
         Path f = dir.resolve("collide.yaml");
         Files.writeString(f, """
@@ -689,32 +685,33 @@ class FleetConfigTest {
                   port: 8080
                 profiles:
                   gx10:
-                    tabLabel: "lead: {profile} #{n}"
+                    tabLabel: "alpha"
                 fleet:
                   leaders:
                     opus:
-                      tab: "lead: opus"
+                      tab: "alpha"
-                      tabPrefix: "lead:"
                 """);
         FleetConfig cfg = FleetConfig.load(f);
 
         IllegalStateException e =
                 assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
         assertTrue(e.getMessage().contains("gx10"), "the message must name the offending profile");
+        assertTrue(e.getMessage().contains("alpha"), "the message must name the offending label");
     }
 
-    /** A bad fleet-wide template promotes every member, not one profile — so it is checked too. */
     @Test
-    void aFleetTabLabelThatMatchesALeadPrefixRefusesToStart(@TempDir Path dir) throws Exception {
+    void aFleetTabLabelTemplateThatCanRenderAsALeadTabRefusesToStart(@TempDir Path dir) throws Exception {
         Path f = dir.resolve("collide-template.yaml");
         Files.writeString(f, """
                 bind:
                   port: 8080
+                profiles:
+                  pha: {}
                 fleet:
-                  tabLabel: "lead: {role} {profile}"
+                  tabLabel: "al{profile}"
                   leaders:
                     opus:
-                      tab: "lead: opus"
+                      tab: "alpha"
                 """);
         FleetConfig cfg = FleetConfig.load(f);
 
@@ -723,12 +720,28 @@ class FleetConfigTest {
         assertTrue(e.getMessage().contains("fleet.tabLabel"));
     }
 
-    /**
-     * The point of making role the label's first field: {@code {role}} comes from a closed enum, so
-     * a generated label cannot begin with {@code "lead:"} however the fleet is configured.
-     */
     @Test
-    void theDefaultTabLabelCannotCollideWithTheDefaultLeadPrefix(@TempDir Path dir) throws Exception {
+    void anExactFleetTabLabelCollisionRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("exact-tab-collision.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  tabLabel: "alpha"
+                  leaders:
+                    alpha:
+                      tab: "alpha"
+                """);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class,
+                () -> FleetConfig.load(f).validateAll());
+        assertTrue(e.getMessage().contains("fleet.tabLabel"),
+                "the message must name the offending label");
+        assertTrue(e.getMessage().contains("alpha"), "the message must name the colliding lead tab");
+    }
+
+    @Test
+    void aFleetTabLabelTemplateThatCannotRenderAsALeadTabIsAllowed(@TempDir Path dir) throws Exception {
         Path f = dir.resolve("ok.yaml");
         Files.writeString(f, """
                 bind:
@@ -737,17 +750,13 @@ class FleetConfigTest {
                   gx10:
                     baseUrl: http://gx00.gw:8000
                 fleet:
+                  tabLabel: "worker-{profile}"
                   leaders:
                     opus:
-                      tab: "lead: opus"
+                      tab: "alpha"
                 """);
 
-        assertDoesNotThrow(() -> FleetConfig.load(f).validateLeadTabPrefixes());
+        assertDoesNotThrow(() -> FleetConfig.load(f).validateAll());
-        for (MemberRole role : MemberRole.values()) {
-            assertFalse(FleetConfig.Fleet.DEFAULT_TAB_LABEL
-                            .replace("{role}", role.wireName()).startsWith("lead:"),
-                    "no role renders a label that reads as a lead");
-        }
     }
 
     @Test
@@ -765,6 +774,78 @@ class FleetConfigTest {
                 "a label that collides with a convention nobody reads is not a problem");
     }
 
+    /**
+     * fleetd #677: identity is matched on a lead's exact {@code tab} alone, so two leads sharing
+     * one tab means only one of them is ever found — the guard must catch this independently of
+     * the member-template checks above.
+     */
+    @Test
+    void twoLeadsSharingTheSameExactTabRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("shared-tab.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "shared tab"
+                    sonnet:
+                      tab: "shared tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
+        assertTrue(e.getMessage().contains("opus"), "the message must name one offending lead");
+        assertTrue(e.getMessage().contains("sonnet"), "the message must name the other offending lead");
+    }
+
+    /**
+     * fleetd #693: the guard matches tabs case-insensitively, because
+     * {@code LeadTabScanner} keys its tab map on a lowercased label — two tabs differing only in
+     * case collide there too, and the guard must catch that independently of the exact-match case
+     * above.
+     */
+    @Test
+    void twoLeadsSharingTheSameTabInDifferentCaseRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("shared-tab-case.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "Shared Tab"
+                    sonnet:
+                      tab: "shared tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
+        assertTrue(e.getMessage().contains("opus"), "the message must name one offending lead");
+        assertTrue(e.getMessage().contains("sonnet"), "the message must name the other offending lead");
+    }
+
+    /** Control for {@link #twoLeadsSharingTheSameExactTabRefusesToStart}: distinct tabs load cleanly. */
+    @Test
+    void twoLeadsWithDistinctExactTabsAreAllowed(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("distinct-tabs.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "opus tab"
+                    sonnet:
+                      tab: "sonnet tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        assertDoesNotThrow(cfg::validateLeadTabPrefixes);
+    }
+
     // ── validatePanePlacementAgainstLeadTabs ────────────────────────────────────────────────────
 
     /**
@@ -919,6 +1000,273 @@ class FleetConfigTest {
                 "no primary.terminal pin ⇒ nothing registered, even with fleet.leaders configured");
     }
 
+    // ── fleetd #669: the collaborators registry ────────────────────────────────────────────────
+
+    /**
+     * {@code Fleet} is {@code @JsonIgnoreProperties(ignoreUnknown = true)}, so a config naming
+     * {@code fleet.collaborators.<name>.tab} loads with no exception whether or not the key is
+     * ever read into the object model. Asserting only "no exception" would pass both before and
+     * after the real fix, so this asserts the parsed value is actually reachable from the loaded
+     * {@code FleetConfig} — the one thing a vacuous "no exception" test cannot tell apart.
+     */
+    @Test
+    void collaboratorsBlockIsActuallyParsedNotSilentlyDropped(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("collaborators.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  collaborators:
+                    reviewer-alex:
+                      tab: "collab: alex"
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertEquals("collab: alex", cfg.fleet().collaborators().get("reviewer-alex").tab());
+    }
+
+    /** A collaborator carries no field other than {@code tab}, so a blank one is meaningless. */
+    @Test
+    void aCollaboratorWithNoTabRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("useless-collaborator.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  collaborators:
+                    ghost: {}
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class, cfg::validateMembers);
+        assertTrue(e.getMessage().contains("ghost"), "the message must name the useless entry");
+        assertTrue(e.getMessage().contains("tab:"), "the message must say what is missing");
+    }
+
+    /** Control for {@link #aCollaboratorWithNoTabRefusesToStart}: a named tab loads cleanly. */
+    @Test
+    void aCollaboratorWithATabIsAllowed(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("named-collaborator.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  collaborators:
+                    reviewer-alex:
+                      tab: "collab: alex"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        assertDoesNotThrow(cfg::validateMembers);
+    }
+
+    /**
+     * fleetd #669: the fleet-wide {@code tabLabel} template can render as a collaborator tab, the
+     * same hazard {@link #aFleetTabLabelTemplateThatCanRenderAsALeadTabRefusesToStart} covers on
+     * the lead side. Drives the fleet-wide branch directly, with no profile override involved.
+     */
+    @Test
+    void aFleetTabLabelTemplateThatCanRenderAsACollaboratorTabRefusesToStart(@TempDir Path dir)
+            throws Exception {
+        Path f = dir.resolve("collide-template-collaborator.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  tabLabel: "al{profile}"
+                  collaborators:
+                    alex:
+                      tab: "alpha"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
+        assertTrue(e.getMessage().contains("fleet.tabLabel"));
+        assertTrue(e.getMessage().contains("alex"), "the message must name the offending collaborator");
+    }
+
+    /**
+     * fleetd #669: a member tabLabel that can render as a configured collaborator tab is the same
+     * hazard as the lead case above — a member labelled that way is read back as the collaborator.
+     */
+    @Test
+    void aProfileTabLabelOverrideMatchingACollaboratorTabRefusesToStart(@TempDir Path dir)
+            throws Exception {
+        Path f = dir.resolve("collide-collaborator.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    tabLabel: "collab-tab"
+                fleet:
+                  collaborators:
+                    alex:
+                      tab: "collab-tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
+        assertTrue(e.getMessage().contains("gx10"), "the message must name the offending profile");
+        assertTrue(e.getMessage().contains("collab-tab"), "the message must name the offending label");
+    }
+
+    /** Control: a profile tabLabel that cannot render as the collaborator tab is allowed. */
+    @Test
+    void aProfileTabLabelThatCannotRenderAsACollaboratorTabIsAllowed(@TempDir Path dir)
+            throws Exception {
+        Path f = dir.resolve("ok-collaborator.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    tabLabel: "worker-{profile}"
+                fleet:
+                  collaborators:
+                    alex:
+                      tab: "collab-tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        assertDoesNotThrow(cfg::validateLeadTabPrefixes);
+    }
+
+    /** fleetd #669: identity is matched on a collaborator's exact tab, so two sharing one are unreachable. */
+    @Test
+    void twoCollaboratorsSharingTheSameExactTabRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("shared-collaborator-tab.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  collaborators:
+                    alex:
+                      tab: "shared tab"
+                    sam:
+                      tab: "Shared Tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
+        assertTrue(e.getMessage().contains("alex"), "the message must name one offending collaborator");
+        assertTrue(e.getMessage().contains("sam"), "the message must name the other offending collaborator");
+    }
+
+    /** Control for {@link #twoCollaboratorsSharingTheSameExactTabRefusesToStart}: distinct tabs load cleanly. */
+    @Test
+    void twoCollaboratorsWithDistinctExactTabsAreAllowed(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("distinct-collaborator-tabs.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  collaborators:
+                    alex:
+                      tab: "alex tab"
+                    sam:
+                      tab: "sam tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        assertDoesNotThrow(cfg::validateLeadTabPrefixes);
+    }
+
+    /**
+     * fleetd #669: a collaborator tab equal to a lead tab crosses a privilege boundary — the worst
+     * of the three new collisions, since only one of the two identities is ever found.
+     */
+    @Test
+    void aCollaboratorTabEqualToALeadTabRefusesToStart(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("lead-collaborator-collision.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "shared tab"
+                  collaborators:
+                    alex:
+                      tab: "Shared Tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, cfg::validateLeadTabPrefixes);
+        assertTrue(e.getMessage().contains("opus"), "the message must name the offending lead");
+        assertTrue(e.getMessage().contains("alex"), "the message must name the offending collaborator");
+    }
+
+    /** Control for {@link #aCollaboratorTabEqualToALeadTabRefusesToStart}: distinct tabs load cleanly. */
+    @Test
+    void aLeadAndACollaboratorWithDistinctTabsAreAllowed(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("lead-collaborator-ok.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  leaders:
+                    opus:
+                      tab: "lead tab"
+                  collaborators:
+                    alex:
+                      tab: "collab tab"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        assertDoesNotThrow(cfg::validateLeadTabPrefixes);
+    }
+
+    /**
+     * fleetd #669: a pane-placed member can land in a collaborator's labelled tab exactly as it
+     * can land in a lead's — {@code validatePanePlacementAgainstLeadTabs()} must fire even when
+     * {@code fleet.leaders} is empty, which is the early-return the brief flagged as the bug.
+     */
+    @Test
+    void aPanePlacedProfileWithACollaboratorTabRefusesToStartEvenWithNoLeaders(@TempDir Path dir)
+            throws Exception {
+        Path f = dir.resolve("pane-hazard-collaborator.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    placement: pane
+                fleet:
+                  collaborators:
+                    alex:
+                      tab: "collab: alex"
+                """);
+        FleetConfig cfg = FleetConfig.load(f);
+
+        IllegalStateException e = assertThrows(IllegalStateException.class,
+                cfg::validatePanePlacementAgainstLeadTabs);
+        assertTrue(e.getMessage().contains("gx10"), "the message must name the offending profile");
+    }
+
+    /** Control: a pane-placed profile with no lead or collaborator tab configured is allowed. */
+    @Test
+    void aPanePlacedProfileWithNoLeaderOrCollaboratorTabIsAllowed(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("pane-no-tab-at-all.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    placement: pane
+                fleet:
+                  collaborators:
+                    alex: {}
+                """);
+
+        assertDoesNotThrow(() -> FleetConfig.load(f).validatePanePlacementAgainstLeadTabs(),
+                "a collaborator with no tab feeds nothing into the scanner, so pane placement is safe");
+    }
+
     // ── CB-548: the architects registry ────────────────────────────────────────────────────────
 
     @Test
@@ -1191,6 +1539,60 @@ class FleetConfigTest {
         assertEquals(Set.of("sonnet"), cfg.fleet().pool(MemberRole.REVIEWER).keySet());
     }
 
+    /**
+     * A duplicated name in {@code fleet.collaborators} is refused at parse time, like any other
+     * {@code fleet:} pool. See {@link #duplicateSlotNamesInOnePoolAreRejectedAtParseTime}.
+     */
+    @Test
+    void duplicateCollaboratorNamesAreRejectedAtParseTime(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("collaborator-dup.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                fleet:
+                  collaborators:
+                    alex:
+                      tab: "collab: alex"
+                    alex:
+                      tab: "collab: alex, second"
+                """);
+
+        IllegalStateException e =
+                assertThrows(IllegalStateException.class, () -> FleetConfig.load(f));
+        assertTrue(e.getMessage().contains("alex"),
+                "the refusal names the duplicated entry, was: " + e.getMessage());
+        assertTrue(e.getMessage().contains("fleet.collaborators"),
+                "the refusal names the pool the duplicate is in, was: " + e.getMessage());
+    }
+
+    /**
+     * Control for {@link #duplicateCollaboratorNamesAreRejectedAtParseTime}: the same name reused
+     * across the collaborators registry and a member role pool is the role × profile matrix doing
+     * its job in the other pool, not a mistake — only a repeat within one pool loses an entry.
+     */
+    @Test
+    void theSameNameInCollaboratorsAndAnotherPoolIsNotADuplicate(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("collaborator-cross-pool.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx10.gw:8000
+                fleet:
+                  architects:
+                    alex:
+                      profile: sonnet
+                  collaborators:
+                    alex:
+                      tab: "collab: alex"
+                """);
+
+        FleetConfig cfg = assertDoesNotThrow(() -> FleetConfig.load(f));
+        assertEquals(Set.of("alex"), cfg.fleet().pool(MemberRole.ARCHITECT).keySet());
+        assertEquals(Set.of("alex"), cfg.fleet().collaborators().keySet());
+    }
+
     @Test
     void duplicateKeysOutsideTheFleetPoolsAreUnaffected(@TempDir Path dir) throws Exception {
         // The duplicate check is scoped to the fleet pools — a duplicate elsewhere is not this
@@ -3175,4 +3577,41 @@ class FleetConfigTest {
         FleetConfig cfg = FleetConfig.load(f);
         assertTrue(cfg.models().offIds().isEmpty());
     }
+
+    // ── fleetd #651: leadRollover.turnSettleSeconds default resolution ─────────────────────────
+
+    @Test
+    void turnSettleSecondsDefaultsTo300WhenUnset(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("bare-rollover.yaml");
+        Files.writeString(f, "bind:\n  port: 8080\nleadRollover: {}\n");
+
+        FleetConfig.LeadRollover rollover = FleetConfig.load(f).leadRollover();
+        assertNotNull(rollover);
+        assertEquals(300, rollover.turnSettleSeconds());
+    }
+
+    @Test
+    void turnSettleSecondsUsesAnExplicitPositiveValue(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("rollover.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                leadRollover:
+                  turnSettleSeconds: 45
+                """);
+
+        FleetConfig.LeadRollover rollover = FleetConfig.load(f).leadRollover();
+        assertEquals(45, rollover.turnSettleSeconds());
+    }
+
+    @Test
+    void turnSettleSecondsFallsBackTo300WhenZeroOrNegative(@TempDir Path dir) throws Exception {
+        Path zero = dir.resolve("zero.yaml");
+        Files.writeString(zero, "bind:\n  port: 8080\nleadRollover:\n  turnSettleSeconds: 0\n");
+        assertEquals(300, FleetConfig.load(zero).leadRollover().turnSettleSeconds());
+
+        Path negative = dir.resolve("negative.yaml");
+        Files.writeString(negative, "bind:\n  port: 8080\nleadRollover:\n  turnSettleSeconds: -5\n");
+        assertEquals(300, FleetConfig.load(negative).leadRollover().turnSettleSeconds());
+    }
 }

fleetd/src/test/java/dev/ltms/fleet/config/FleetConfigValidateAllTest.java

@@ -17,63 +17,21 @@ import static org.junit.jupiter.api.Assertions.assertThrows;
 import static org.junit.jupiter.api.Assertions.assertTrue;
 
 /**
- * The gap this class exists to close: mutation testing on the fleetd ticket "central allow-list
+ * Tests the reflective validator sweep and {@link FleetConfig#validateAll()} reachability.
- * of usable models" found that although {@link FleetConfig#validateModels()}'s own logic was well
- * pinned, nothing proved either real caller ({@code Fleetd.main} and {@link ConfigRef#reload()})
- * still invoked it — deleting the call site left the full suite green (1478/0/0/0). A follow-up
- * measurement (same technique — remove one call site, run the suite, not read the code) found the
- * SAME gap for all five of {@link FleetConfig}'s other validators at startup, and for four of the
- * six inside {@link ConfigRef#reload()}. This is a class of gap, not one line's mistake: every one
- * of those thirteen tests called the validator itself directly, never the real caller that was
- * supposed to.
  *
- * <p>The fix replaces the six individual {@code cfg.validateXxx()} calls at each of the two real
+ * <p>{@link #theSweepRunsEveryValidateMethodOnAnUnrelatedClass()} and its neighbours
- * call sites with one {@link FleetConfig#validateAll()}, which reaches every validator by
+ * prove that {@link FleetConfig#invokeAllValidators} runs each public, no-arg, void
- * reflection rather than by a hand-maintained list of names. A hand-maintained list of six names
+ * {@code validateXxx()} method on its target. {@link #fleetConfigDeclaresExactlyTheseValidatorsToday()}
- * would have exactly the defect it replaces: the seventh validator someone adds next month has no
+ * is the canary for the validator set. {@link #validateAllReachesEveryOneOfTodaysRealValidators()}
- * reason to be added to it, and nothing would say so. This class proves TWO separate claims, and
+ * is the reachability check for that set.
- * keeps them separate on purpose:
- *
- * <ol>
- *   <li>{@link #theSweepMechanismIsGenericNotHardcodedToFleetConfigsSixNames()} and its neighbours
- *       prove the reflective sweep itself ({@link FleetConfig#invokeAllValidators}) is a general
- *       mechanism — it runs whatever public, no-arg, void {@code validateXxx()} methods a class
- *       happens to declare today, including a class with more of them than {@link FleetConfig}
- *       has right now. This is the proof that a future, real seventh validator on {@link
- *       FleetConfig} would be swept automatically, without needing to add a real (unwanted)
- *       seventh validator just to exercise the claim.</li>
- *   <li>{@link #validateAllReachesEveryOneOfTodaysRealValidators()} proves {@link
- *       FleetConfig#validateAll()} itself is wired to that same generic mechanism and genuinely
- *       reaches seven of today's eight real validators — reusing the exact minimal failing
- *       configurations {@code FleetConfigTest} already established for each one directly, so a
- *       single call to {@code validateAll()} is shown to reproduce every one of those seven
- *       failures. The eighth, {@link FleetConfig#validateLeadRollover()}, has no case here yet —
- *       a pre-existing gap tracked as fleetd #668.</li>
- * </ol>
- *
- * <p>Together with the direct-{@code Fleetd.main}-invocation tests in {@code
- * FleetdStartupValidationTest} (which prove the real startup call site still calls {@code
- * validateAll()}) and the {@code ConfigRefTest} reload tests (which prove the same for {@link
- * ConfigRef#reload()}), removing {@code cfg.validateAll();} from either real call site now fails
- * a test in this module.
- *
- * <p><b>What is NOT pinned, measured rather than assumed.</b> Reverting {@link
- * FleetConfig#validateAll()} to a hardcoded list of today's six method calls leaves the whole
- * suite green (measured at review: 1491 tests, 0 failures). Nothing ties {@code validateAll()} to
- * the generic sweep — claim 1 proves {@link FleetConfig#invokeAllValidators} is generic, and claim
- * 2 proves {@code validateAll()} reaches today's six, and a hardcoded list satisfies both. So the
- * reflective sweep is a convenience, not the guarantee. The guarantee is {@link
- * #fleetConfigDeclaresExactlyTheseValidatorsToday()}: it fails the moment any validator is added
- * or removed, which forces whoever changes the set to look at this file.
  */
 class FleetConfigValidateAllTest {
 
-    // ── Claim 1: the reflective sweep is a general mechanism, not six names in disguise ──────────
+    // ── Claim 1: the reflective sweep is a general mechanism ─────────────────────────────────────
 
     /**
-     * A throwaway fixture class, unrelated to {@link FleetConfig} in every way except shape: three
+     * Fixture with public, no-arg, void methods named {@code validateXxx}. It proves the sweep uses
-     * public, no-arg, void methods named {@code validateXxx}. Proves the sweep works on ANY class
+     * the target's method shape rather than special handling for {@link FleetConfig}.
-     * with this shape, not on something special-cased to {@link FleetConfig}.
      */
     static class ThreeValidators {
         final List<String> ran = new ArrayList<>();
@@ -92,7 +50,7 @@ class FleetConfigValidateAllTest {
     }
 
     @Test
-    void theSweepMechanismIsGenericNotHardcodedToFleetConfigsSixNames() {
+    void theSweepRunsEveryValidateMethodOnAnUnrelatedClass() {
         ThreeValidators target = new ThreeValidators();
         FleetConfig.invokeAllValidators(target);
         assertEquals(List.of("validateAlpha", "validateBeta", "validateGamma"), target.ran,
@@ -102,12 +60,8 @@ class FleetConfigValidateAllTest {
     }
 
     /**
-     * The core of the "self-maintaining" requirement: the exact same class shape as {@link
+     * Fixture with an added valid method. It proves the sweep reaches a method because it matches
-     * ThreeValidators}, plus one more method — standing in for "a developer adds a validator next
+     * the validator shape.
-     * month". Nothing about the sweep changes to pick it up; the new method is invoked purely
-     * because it exists and matches the shape. This is what makes adding a seventh real validator
-     * to {@link FleetConfig} safe without touching {@link FleetConfig#validateAll()} or either
-     * call site — there is no "wire it in" step left to forget.
      */
     static class FourValidators {
         final List<String> ran = new ArrayList<>();
@@ -135,7 +89,7 @@ class FleetConfigValidateAllTest {
         FleetConfig.invokeAllValidators(target);
         assertEquals(List.of("validateAlpha", "validateBeta", "validateDelta", "validateGamma"),
                 sorted(target.ran),
-                "the fourth method must be reached automatically — proving a class can grow the "
+                "the added method must be reached automatically — proving a class can grow the "
                         + "set of things it validates with no change to the sweep itself");
     }
 
@@ -210,7 +164,7 @@ class FleetConfigValidateAllTest {
                         + "name) must all be skipped");
     }
 
-    // ── Claim 2: FleetConfig.validateAll() is wired to that mechanism and reaches seven of eight today ──
+    // ── Claim 2: FleetConfig.validateAll() is wired to that mechanism and reaches every validator ──
 
     /**
      * Reflectively enumerates {@link FleetConfig}'s own public, no-arg, void {@code validateXxx()}
@@ -220,6 +174,11 @@ class FleetConfigValidateAllTest {
      * the {@code Set.of} below, so a reader adding or removing one sees this assertion name the new
      * count rather than a silent pass at the old one. The count lives only in that set, not in this
      * method's name, so the two cannot drift apart.
+     *
+     * <p>This assertion alone proves only that the validator exists with the right shape — it
+     * cannot prove {@code validateAll()} actually reaches it. Only {@link
+     * #validateAllReachesEveryOneOfTodaysRealValidators()} proves reachability, which is why this
+     * method's failure message sends the reader there too.
      */
     @Test
     void fleetConfigDeclaresExactlyTheseValidatorsToday() {
@@ -237,11 +196,16 @@ class FleetConfigValidateAllTest {
                 "validateSubscriptionProfiles", "validateCharters", "validateMembers",
                 "validateModels", "validateLeadRollover", "validatePanePlacementAgainstLeadTabs")),
                 names,
-                "FleetConfig's public validate*() methods changed. Do TWO things, in this "
+                "FleetConfig's public validate*() methods changed. Do THREE things, in this "
                         + "order. First confirm validateAll() still delegates to "
                         + "invokeAllValidators(this) — a hardcoded list there passes every other "
                         + "test in this class, so this assertion is the only place that will ever "
-                        + "make you check. Only then update the expected set to match.");
+                        + "make you check. Second, update the expected set below to match. Third, "
+                        + "add or remove a case for that validator in "
+                        + "validateAllReachesEveryOneOfTodaysRealValidators() below — this "
+                        + "assertion proves only that the validator exists with the right shape, "
+                        + "never that validateAll() reaches it; that enumeration is the test that "
+                        + "does.");
     }
 
     /** A minimal, otherwise-valid file — same shape FleetConfigTest and ConfigRefTest use. */
@@ -265,14 +229,18 @@ class FleetConfigValidateAllTest {
     }
 
     /**
-     * The heart of claim 2: for seven of today's eight real validators, a minimal file that fails
+     * The heart of claim 2: for every one of today's real validators, a minimal file that
-     * ONLY that one — the exact fixtures {@code FleetConfigTest} uses to test each validator
+     * fails ONLY that one — the exact fixtures {@code FleetConfigTest} uses to test each validator
-     * directly — must also fail through {@link FleetConfig#validateAll()}. If a future edit to
+     * directly, or a dedicated minimal fixture where no other test drives that validator through
-     * {@code validateAll()} silently dropped one of these seven from the sweep (e.g. a typo'd name
+     * {@code validateAll()} — must also fail through {@link FleetConfig#validateAll()}. If a
-     * filter), exactly one of them would start passing when it must not.
+     * future edit to {@code validateAll()} silently dropped one of these from the sweep
+     * (e.g. a typo'd name filter), exactly one of them would start passing when it must not.
      *
-     * <p>The eighth, {@link FleetConfig#validateLeadRollover()}, has no case here — a pre-existing
+     * <p>This is the single place that proves {@code validateAll()} reaches a given validator.
-     * gap tracked as fleetd #668, not fixed by this change.
+     * Adding or removing a validator on {@link FleetConfig} must add or remove a case here, not
+     * only an updated name in {@link #fleetConfigDeclaresExactlyTheseValidatorsToday()}'s expected
+     * set — that assertion proves the validator's shape, never that {@code validateAll()} reaches
+     * it.
      */
     @Test
     void validateAllReachesEveryOneOfTodaysRealValidators(@TempDir Path dir) throws Exception {
@@ -283,16 +251,16 @@ class FleetConfigValidateAllTest {
                   port: 8765
                 """, "auth.mode: token");
 
-        // validateLeadTabPrefixes: a fleet-wide tabLabel that starts with a lead's own tabPrefix.
+        // validateLeadTabPrefixes: a fleet-wide tabLabel that equals a lead tab.
         assertValidateAllRefuses(dir, "lead-tab-prefixes.yaml", """
                 bind:
                   host: 127.0.0.1
                   port: 8765
                 fleet:
-                  tabLabel: "lead: {role} {profile}"
+                  tabLabel: "alpha"
                   leaders:
                     opus:
-                      tab: "lead: opus"
+                      tab: "alpha"
                 """, "fleet.tabLabel");
 
         // validateSubscriptionProfiles: subscription: true with env: reseating ANTHROPIC_BASE_URL.
@@ -362,6 +330,15 @@ class FleetConfigValidateAllTest {
                     opus:
                       tab: "lead: opus"
                 """, "gx10");
+
+        // validateLeadRollover: a leadRollover: block present with no handoverPath.
+        assertValidateAllRefuses(dir, "lead-rollover.yaml", """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                leadRollover:
+                  requireOperatorConfirm: false
+                """, "handoverPath");
     }
 
     private static void assertValidateAllRefuses(Path dir, String fileName, String yaml,

fleetd/src/test/java/dev/ltms/fleet/lead/LeadContextGaugeHighThresholdTest.java

@@ -81,12 +81,6 @@ class LeadContextGaugeHighThresholdTest {
         assertEquals(LeadContextGauge.State.HIGH,
                 atGauge.read(atConfigDir, SESSION_ID, "claude", null).state(),
                 "the fixed default must still be 200,000 when no window is resolvable");
-
-        String legacyConfigDir = writeTranscript(tmp.resolve("legacy"), SESSION_ID, 200_000);
-        LeadContextGauge legacyGauge = new LeadContextGauge();
-        assertEquals(LeadContextGauge.State.HIGH,
-                legacyGauge.read(legacyConfigDir, SESSION_ID, "claude").state(),
-                "the 3-arg read() (no window argument at all) must behave exactly like passing a null window");
     }
 
     @Test

fleetd/src/test/java/dev/ltms/fleet/lead/LeadContextGaugeTest.java

@@ -85,7 +85,7 @@ class LeadContextGaugeTest {
                 usageLine(40_000, 5_000, 3_000)); // last record: 48,000
         LeadContextGauge gauge = new LeadContextGauge();
 
-        LeadContextGauge.Reading first = gauge.read(configDir, SESSION_ID, "claude");
+        LeadContextGauge.Reading first = gauge.read(configDir, SESSION_ID, "claude", null);
         assertEquals(48_000L, first.tokens(), "must total input+cache_read+cache_creation of the LAST usage record");
         assertEquals(LeadContextGauge.State.OK, first.state());
 
@@ -93,7 +93,7 @@ class LeadContextGaugeTest {
         // must change with it, not stay pinned to the first fixture's total.
         String otherSession = "22222222-2222-2222-2222-222222222222";
         writeTranscript(tmp, otherSession, usageLine(100_000, 50_000, 50_000)); // last record: 200,000
-        LeadContextGauge.Reading second = gauge.read(configDir, otherSession, "claude");
+        LeadContextGauge.Reading second = gauge.read(configDir, otherSession, "claude", null);
         assertEquals(200_000L, second.tokens());
         assertTrue(second.tokens() != first.tokens(), "changing N in the fixture must change the reported number");
     }
@@ -111,12 +111,12 @@ class LeadContextGaugeTest {
                 compactionLine(),
                 usageLine(3_000, 0, 0));
         LeadContextGauge gauge = new LeadContextGauge();
-        LeadContextGauge.Reading twoCompactions = gauge.read(tmp.toString(), sessionTwoCompactions, "claude");
+        LeadContextGauge.Reading twoCompactions = gauge.read(tmp.toString(), sessionTwoCompactions, "claude", null);
         assertEquals(2, twoCompactions.compactions());
 
         String sessionZeroCompactions = "44444444-4444-4444-4444-444444444444";
         writeTranscript(tmp, sessionZeroCompactions, usageLine(3_000, 0, 0));
-        LeadContextGauge.Reading zeroCompactions = gauge.read(tmp.toString(), sessionZeroCompactions, "claude");
+        LeadContextGauge.Reading zeroCompactions = gauge.read(tmp.toString(), sessionZeroCompactions, "claude", null);
         assertEquals(0, zeroCompactions.compactions(), "changing K in the fixture must change the reported count");
     }
 
@@ -126,7 +126,7 @@ class LeadContextGaugeTest {
     @DisplayName("a missing transcript file reports UNKNOWN with no token number")
     void missingFileIsUnknown(@TempDir Path tmp) {
         LeadContextGauge gauge = new LeadContextGauge();
-        LeadContextGauge.Reading reading = gauge.read(tmp.toString(), SESSION_ID, "claude");
+        LeadContextGauge.Reading reading = gauge.read(tmp.toString(), SESSION_ID, "claude", null);
         assertEquals(LeadContextGauge.State.UNKNOWN, reading.state());
         assertNull(reading.tokens());
     }
@@ -148,7 +148,7 @@ class LeadContextGaugeTest {
             assumeFalse(Files.isReadable(file),
                     "runs as root (CI container): the read bit does not stop root, so this case cannot be set up here");
             LeadContextGauge gauge = new LeadContextGauge();
-            LeadContextGauge.Reading reading = gauge.read(configDir, SESSION_ID, "claude");
+            LeadContextGauge.Reading reading = gauge.read(configDir, SESSION_ID, "claude", null);
             assertEquals(LeadContextGauge.State.UNKNOWN, reading.state());
             assertNull(reading.tokens());
         } finally {
@@ -171,7 +171,7 @@ class LeadContextGaugeTest {
         Files.writeString(file, lastCompleteLine + "\n" + tornLine, StandardCharsets.UTF_8);
 
         LeadContextGauge gauge = new LeadContextGauge();
-        LeadContextGauge.Reading reading = gauge.read(tmp.toString(), SESSION_ID, "claude");
+        LeadContextGauge.Reading reading = gauge.read(tmp.toString(), SESSION_ID, "claude", null);
         assertEquals(LeadContextGauge.State.OK, reading.state(),
                 "a torn final line must not turn a good earlier reading into UNKNOWN");
         assertEquals(6_000L, reading.tokens(),
@@ -187,7 +187,7 @@ class LeadContextGaugeTest {
                 "{this is not json at all",
                 "neither is this{{{");
         LeadContextGauge gauge = new LeadContextGauge();
-        LeadContextGauge.Reading reading = gauge.read(configDir, SESSION_ID, "claude");
+        LeadContextGauge.Reading reading = gauge.read(configDir, SESSION_ID, "claude", null);
         assertEquals(LeadContextGauge.State.UNKNOWN, reading.state(),
                 "every line unparseable is the real format-change signal and must still report UNKNOWN");
         assertNull(reading.tokens());
@@ -224,12 +224,12 @@ class LeadContextGaugeTest {
         AtomicLong now = new AtomicLong(0);
         LeadContextGauge gauge = new LeadContextGauge(now::get, 5_000);
 
-        gauge.read(configDir, SESSION_ID, "claude");
+        gauge.read(configDir, SESSION_ID, "claude", null);
-        gauge.read(configDir, SESSION_ID, "claude"); // still inside the TTL window
+        gauge.read(configDir, SESSION_ID, "claude", null); // still inside the TTL window
         assertEquals(1, gauge.diskReadCount(), "two reads inside the TTL must touch disk once");
 
         now.set(6_000); // past the TTL
-        gauge.read(configDir, SESSION_ID, "claude");
+        gauge.read(configDir, SESSION_ID, "claude", null);
         assertEquals(2, gauge.diskReadCount(), "a read past the TTL must touch disk again");
     }
 
@@ -241,8 +241,8 @@ class LeadContextGaugeTest {
         String configDir = writeTranscript(tmp, SESSION_ID, usageLine(1_000, 0, 0));
         LeadContextGauge gauge = new LeadContextGauge();
 
-        assertEquals(LeadContextGauge.State.UNKNOWN, gauge.read(configDir, SESSION_ID, "opencode").state());
+        assertEquals(LeadContextGauge.State.UNKNOWN, gauge.read(configDir, SESSION_ID, "opencode", null).state());
-        assertEquals(LeadContextGauge.State.UNKNOWN, gauge.read(configDir, SESSION_ID, null).state());
+        assertEquals(LeadContextGauge.State.UNKNOWN, gauge.read(configDir, SESSION_ID, null, null).state());
-        assertEquals(LeadContextGauge.State.UNKNOWN, gauge.read(configDir, null, "claude").state());
+        assertEquals(LeadContextGauge.State.UNKNOWN, gauge.read(configDir, null, "claude", null).state());
     }
 }

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

@@ -97,6 +97,7 @@ class FleetMcpAuthzTest {
     private static final Principal WORKER_A = Principal.worker("term_a", 200);
     private static final Principal ANON = Principal.anonymous();
     private static final Principal ARCH_DESIGN = Principal.architect("lead-designer", "term_design", 400);
+    private static final Principal COLLABORATOR = Principal.collaborator("ops", "term_collab", 600);
 
     // --- the table, enforced on THIS path too ---------------------------------------------------
 
@@ -156,6 +157,45 @@ class FleetMcpAuthzTest {
         }
     }
 
+    /**
+     * fleetd #669 Unit A: {@code SEND} is split into three actions ({@link Authz.Action#SEND},
+     * {@link Authz.Action#COORD_SEND}, {@link Authz.Action#ANSWER}), each carrying the same grant
+     * the one undivided action gave. An architect holds all three, exactly as it held the one.
+     */
+    @Test
+    void anArchitectMayUseAllThreeSendShapesOverMcp() {
+        FleetMcp m = mcp(true);
+        for (Authz.Action a : new Authz.Action[]{Authz.Action.SEND, Authz.Action.COORD_SEND,
+                Authz.Action.ANSWER}) {
+            assertNull(m.denyFor(ARCH_DESIGN, a, "term_a"),
+                    a + " carries the same grant the undivided SEND action gave an architect");
+        }
+    }
+
+    /** The other half of the same split: a worker is excluded from all three, as it was from one. */
+    @Test
+    void aWorkerMayNotUseAnySendShapeOverMcp() {
+        FleetMcp m = mcp(true);
+        for (Authz.Action a : new Authz.Action[]{Authz.Action.SEND, Authz.Action.COORD_SEND,
+                Authz.Action.ANSWER}) {
+            McpSchema.CallToolResult denied = m.denyFor(WORKER_A, a, "term_a");
+            assertNotNull(denied, a + " must stay refused to a worker");
+            assertTrue(denied.isError(), "a refusal is returned as an MCP tool error");
+        }
+    }
+
+    /**
+     * fleetd #669 Unit A / #678: {@code TASK_READ} (ticket polling, session status) is split out of
+     * the roster-only {@code READ}, carrying forward the grant the undivided action gave. A worker
+     * still has both — it never gained or lost anything by the split.
+     */
+    @Test
+    void aWorkerKeepsBothReadActionsAfterTheSplit() {
+        FleetMcp m = mcp(true);
+        assertNull(m.denyFor(WORKER_A, Authz.Action.READ, null));
+        assertNull(m.denyFor(WORKER_A, Authz.Action.TASK_READ, null));
+    }
+
     @Test
     void anArchitectMayReplyAndAskOnlyAsItsOwnPaneOverMcp() {
         FleetMcp m = mcp(true);
@@ -187,6 +227,19 @@ class FleetMcpAuthzTest {
                 "the caller IS authenticated — it is just not the right role");
     }
 
+    /**
+     * {@code denyFor} passes the real production classifier, not a test-supplied one — no
+     * terminal is recognised as a configured lead or collaborator, so a collaborator's SEND is
+     * refused over MCP.
+     */
+    @Test
+    void aCollaboratorMayNotSendOverMcpWithTheRealProductionClassifier() {
+        FleetMcp m = mcp(true);
+        McpSchema.CallToolResult denied = m.denyFor(COLLABORATOR, Authz.Action.SEND, "term_lead");
+        assertNotNull(denied, "no terminal is recognised as a lead or collaborator yet");
+        assertTrue(denied.isError());
+    }
+
     @Test
     void theLegacyConstructorLeavesTheGateOpen() {
         // The 22 pre-existing FleetMcpTest cases rely on no authorization being enforced.
@@ -297,35 +350,53 @@ class FleetMcpAuthzTest {
      * the whole time the defect was live -- the table was right, the action fed to it was wrong.
      */
     @Test
-    void pollingByTargetIsADrainAndPollingByTicketIsARead() {
+    void pollingByTargetIsADrainAndPollingByTicketIsATaskRead() {
         assertEquals(Authz.Action.DRAIN, FleetMcp.pollAction("term_b", null),
                 "poll by target removes the replies — that is a drain, not an observation");
-        assertEquals(Authz.Action.READ, FleetMcp.pollAction(null, null),
+        assertEquals(Authz.Action.TASK_READ, FleetMcp.pollAction(null, null),
-                "poll by ticket changes nothing");
+                "poll by ticket changes nothing, but is not the roster-only READ action");
-        assertEquals(Authz.Action.READ, FleetMcp.pollAction("  ", null),
+        assertEquals(Authz.Action.TASK_READ, FleetMcp.pollAction("  ", null),
                 "a blank target is an absent target");
     }
 
     /**
      * fleetd #421: a coordId branch is a THIRD operation behind fleet_poll's one name, and it must
-     * map to {@link Authz.Action#COORD_READ} — never {@link Authz.Action#READ}, even though this
+     * map to {@link Authz.Action#COORD_READ} — never {@link Authz.Action#READ} or {@link
-     * branch also consumes nothing. READ's grant is open to every authenticated role on the premise
+     * Authz.Action#TASK_READ}, even though this branch also consumes nothing. A lead-to-lead body
-     * that the roster carries no secrets; a lead-to-lead body is not the roster, so folding this
+     * is not the roster and not a ticket/status read, so folding this branch into either would let
-     * branch into READ would let any worker read every peer lead's mail in full. coordId also takes
+     * any worker or architect read every peer lead's mail in full. coordId also takes priority over
-     * priority over target when both happen to be present — it addresses a different inbox entirely.
+     * target when both happen to be present — it addresses a different inbox entirely.
      */
     @Test
-    void pollingByCoordIdIsACoordReadNeverAPlainRead() {
+    void pollingByCoordIdIsACoordReadNeverAPlainOrTaskRead() {
         assertEquals(Authz.Action.COORD_READ, FleetMcp.pollAction(null, "mac-opus"),
-                "reading held peer mail must not be mapped to the everyone-readable READ action");
+                "reading held peer mail must not be mapped to a widely-readable action");
         assertEquals(Authz.Action.COORD_READ, FleetMcp.pollAction("  ", "mac-opus"),
                 "a blank target must not fall through to READ/DRAIN when coordId is present");
-        assertEquals(Authz.Action.READ, FleetMcp.pollAction(null, "  "),
+        assertEquals(Authz.Action.TASK_READ, FleetMcp.pollAction(null, "  "),
                 "a blank coordId is an absent coordId, same as target/ticket");
         assertEquals(Authz.Action.COORD_READ, FleetMcp.pollAction("term_b", "mac-opus"),
                 "coordId takes priority over target — this is a different inbox, not a drain");
     }
 
+    /**
+     * {@code fleet_send} is three call shapes behind one tool name, exactly as {@code fleet_poll}
+     * is (fleetd #669 Unit A). {@link FleetMcp#sendAction} picks the action from the arguments, not
+     * the handler, for the same reason {@link FleetMcp#pollAction} does: a test can assert the
+     * mapping the handler actually uses.
+     */
+    @Test
+    void sendMapsToThreeDifferentActionsByItsArguments() {
+        assertEquals(Authz.Action.SEND, FleetMcp.sendAction(null, null),
+                "a plain delivery, with neither coordId nor turnId, is a local SEND");
+        assertEquals(Authz.Action.COORD_SEND, FleetMcp.sendAction("mac-opus", null),
+                "coordId addresses a peer lead over the coordination broker");
+        assertEquals(Authz.Action.ANSWER, FleetMcp.sendAction(null, "turn-1"),
+                "turnId resolves a worker's blocked question");
+        assertEquals(Authz.Action.COORD_SEND, FleetMcp.sendAction("mac-opus", "turn-1"),
+                "coordId takes priority over turnId, mirroring sendToLead's own mutual-exclusion check");
+    }
+
     @Test
     void everyRegisteredToolHasItsHandlerActionPinned() {
         // fleetd #469: this used to scrape FleetMcp.java's tool("…") calls for the registered set —
@@ -344,16 +415,22 @@ class FleetMcpAuthzTest {
                 () -> tool + " is registered but has no pinned authorization action"));
 
         assertEquals(Authz.Action.SEND, FleetMcp.toolAction("fleet_send", Map.of()));
+        assertEquals(Authz.Action.SEND,
+                FleetMcp.toolAction("fleet_send", Map.of("sessionId", "term_a", "content", "hi")));
+        assertEquals(Authz.Action.COORD_SEND,
+                FleetMcp.toolAction("fleet_send", Map.of("coordId", "mac-opus", "content", "hi")));
+        assertEquals(Authz.Action.ANSWER,
+                FleetMcp.toolAction("fleet_send", Map.of("turnId", "turn-1", "content", "hi")));
         assertEquals(Authz.Action.REPLY, FleetMcp.toolAction("fleet_reply", Map.of()));
         assertEquals(Authz.Action.ASK, FleetMcp.toolAction("fleet_ask", Map.of()));
-        assertEquals(Authz.Action.READ, FleetMcp.toolAction("fleet_status", Map.of()));
+        assertEquals(Authz.Action.TASK_READ, FleetMcp.toolAction("fleet_status", Map.of()));
         assertEquals(Authz.Action.DRAIN, FleetMcp.toolAction("fleet_ack", Map.of()));
         assertEquals(Authz.Action.SPAWN, FleetMcp.toolAction("fleet_spawn", Map.of()));
         assertEquals(Authz.Action.READ, FleetMcp.toolAction("fleet_list", Map.of()));
         assertEquals(Authz.Action.STOP, FleetMcp.toolAction("fleet_stop", Map.of()));
         assertEquals(Authz.Action.READ, FleetMcp.toolAction("fleet_profiles", Map.of()));
         assertEquals(Authz.Action.READ, FleetMcp.toolAction("fleet_whoami", Map.of()));
-        assertEquals(Authz.Action.READ, FleetMcp.toolAction("fleet_poll", Map.of("ticket", "task")));
+        assertEquals(Authz.Action.TASK_READ, FleetMcp.toolAction("fleet_poll", Map.of("ticket", "task")));
         assertEquals(Authz.Action.DRAIN, FleetMcp.toolAction("fleet_poll", Map.of("target", "term_b")));
         assertEquals(Authz.Action.COORD_READ,
                 FleetMcp.toolAction("fleet_poll", Map.of("coordId", "mac-opus")));

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

@@ -1837,6 +1837,32 @@ class FleetMcpTest {
         assertTrue(out.contains("\"sessionId\":\"term_design\""), out);
     }
 
+    /**
+     * A collaborator reports its own role and name, never the {@code leader} key a lead gets —
+     * {@code role} already reads {@code "collaborator"}, so a {@code leader} key alongside it
+     * would be self-contradicting. Control: the same call shape fed a named lead must still carry
+     * {@code leader}, so this is not passing because the key stopped being emitted for everyone.
+     */
+    @Test
+    void whoamiReportsACollaboratorWithNoLeaderKeyButALeadStillGetsOne() {
+        FakeHerdr h = new FakeHerdr();
+        SessionManager sessions = sessionManager(h, "http://gx00.gw:8000", Set.of("gx00.gw"));
+
+        McpSchema.CallToolResult collabRes = FleetMcp.whoami(
+                Principal.collaborator("ops", "term_collab", 700), sessions);
+        assertNotEquals(Boolean.TRUE, collabRes.isError());
+        String collabOut = textOf(collabRes);
+        assertTrue(collabOut.contains("\"role\":\"collaborator\""), collabOut);
+        assertTrue(collabOut.contains("\"collaborator\":\"ops\""), collabOut);
+        assertTrue(collabOut.contains("\"sessionId\":\"term_collab\""), collabOut);
+        assertFalse(collabOut.contains("leader"), collabOut);
+
+        McpSchema.CallToolResult leadRes = FleetMcp.whoami(
+                Principal.leader("opus", "term_lead", 100), sessions);
+        String leadOut = textOf(leadRes);
+        assertTrue(leadOut.contains("\"leader\":\"opus\""), leadOut);
+    }
+
     /**
      * CB-548: an architect SEND delegates as its own pane (recording the per-target delegation) but
      * must NEVER become the legacy singleton "primary" fallback — the per-target map does not cure

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

@@ -60,13 +60,25 @@ class MessageServiceTest {
         inbox.own(T);
     }
 
+    /**
+     * A send budget large enough that a test's own setup — {@link #awaitWaiting()} plus whatever
+     * status transitions it drives afterward — can never compete with it for the same clock. A test
+     * that needs {@code send.get(...)}'s own window to be the only timing bound it depends on uses
+     * {@link #sendAsync(String, long)} with this value instead of the default 5000 ms.
+     */
+    private static final long GENEROUS_SEND_BUDGET_MILLIS = 30_000;
+
     /** Run {@code send} on a background thread; the current thread drives the worker's turn. */
     private CompletableFuture<MessageService.Reply> sendAsync() {
         return sendAsync("do the task");
     }
 
     private CompletableFuture<MessageService.Reply> sendAsync(String content) {
-        return CompletableFuture.supplyAsync(() -> messages.send(T, content, 5000));
+        return sendAsync(content, 5000);
+    }
+
+    private CompletableFuture<MessageService.Reply> sendAsync(String content, long timeoutMillis) {
+        return CompletableFuture.supplyAsync(() -> messages.send(T, content, timeoutMillis));
     }
 
     private void awaitWaiting() throws InterruptedException {
@@ -80,7 +92,7 @@ class MessageServiceTest {
 
     @Test
     void completionFallbackResolvesATurnThatNeverCalledFleetReply() throws Exception {
-        CompletableFuture<MessageService.Reply> send = sendAsync();
+        CompletableFuture<MessageService.Reply> send = sendAsync("do the task", GENEROUS_SEND_BUDGET_MILLIS);
         awaitWaiting();
 
         herdr.readText("$ prompt");                // pre-turn pane: no answer yet (baseline reference)
@@ -96,6 +108,32 @@ class MessageServiceTest {
         assertTrue(reply.completed(), "a scraped completion still counts as completed");
     }
 
+    /**
+     * Pins {@link #GENEROUS_SEND_BUDGET_MILLIS} as the budget {@link
+     * #completionFallbackResolvesATurnThatNeverCalledFleetReply} depends on. A 5500 ms delay between
+     * {@link #awaitWaiting()} and the status transitions that drive completion stands in for a loaded
+     * machine's setup overhead — comfortably past the 5000 ms budget this send no longer uses, and
+     * still well inside this method's own 30 000 ms budget. The only clock this test depends on is
+     * {@code send.get}'s own 10 s window.
+     */
+    @Test
+    void completionFallbackSurvivesASlowHarnessBecauseItsSendBudgetIsNotTheBindingClock() throws Exception {
+        CompletableFuture<MessageService.Reply> send = sendAsync("do the task", GENEROUS_SEND_BUDGET_MILLIS);
+        awaitWaiting();
+
+        Thread.sleep(5500);
+
+        herdr.readText("$ prompt");
+        injector.onStatus(T, AgentStatus.IDLE);
+        injector.onStatus(T, AgentStatus.WORKING);
+        herdr.readText("BUILD GREEN: 391 files");
+        injector.onStatus(T, AgentStatus.IDLE);
+
+        MessageService.Reply reply = send.get(10, TimeUnit.SECONDS);
+        assertEquals(MessageService.Outcome.COMPLETED_UNREPLIED, reply.outcome(),
+                "a slow harness must not be mistaken for a timed-out delivery");
+    }
+
     @Test
     void completionFallbackReplacesAnEchoedInjectedBriefWithNoReportOutcome() throws Exception {
         String brief = "Implement the requested change. ".repeat(20);

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

@@ -1,8 +1,12 @@
 package dev.ltms.fleet.rest;
 
+import ch.qos.logback.classic.Level;
+import ch.qos.logback.classic.spi.ILoggingEvent;
+import com.fasterxml.jackson.databind.ObjectMapper;
 import dev.ltms.fleet.auth.CallerResolver;
 import dev.ltms.fleet.auth.Authz;
 import dev.ltms.fleet.auth.MemberRegistry;
+import dev.ltms.fleet.auth.Principal;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.guard.SubscriptionGuard;
 import dev.ltms.fleet.herdr.AgentControl;
@@ -18,6 +22,7 @@ import dev.ltms.fleet.msg.Rendezvous;
 import dev.ltms.fleet.session.FakeWorktrees;
 import dev.ltms.fleet.session.SessionManager;
 import dev.ltms.fleet.member.ClaudeCodeLauncher;
+import dev.ltms.fleet.testing.CapturedLog;
 import io.javalin.Javalin;
 import org.junit.jupiter.api.AfterEach;
 import org.junit.jupiter.api.Test;
@@ -28,10 +33,13 @@ import java.net.http.HttpRequest;
 import java.net.http.HttpResponse;
 import java.nio.file.Files;
 import java.nio.file.Path;
+import java.util.ArrayList;
 import java.util.LinkedHashSet;
+import java.util.List;
 import java.util.Locale;
 import java.util.Map;
 import java.util.Set;
+import java.util.function.Predicate;
 import java.util.regex.Matcher;
 import java.util.regex.Pattern;
 import java.util.stream.Collectors;
@@ -122,12 +130,60 @@ class FleetAppAuthTest {
         assertEquals(Authz.Action.DRAIN, FleetApp.routeAction("GET /sessions/{id}/replies"));
         assertEquals(Authz.Action.ASK, FleetApp.routeAction("POST /sessions/{id}/ask"));
         for (String route : Set.of("GET /sessions", "GET /agents", "GET /members", "GET /profiles",
-                "GET /member-credentials", "GET /sessions/{id}/status", "GET /tasks/{ticket}")) {
+                "GET /member-credentials")) {
             assertEquals(Authz.Action.READ, FleetApp.routeAction(route), route);
         }
+        for (String route : Set.of("GET /sessions/{id}/status", "GET /tasks/{ticket}")) {
+            assertEquals(Authz.Action.TASK_READ, FleetApp.routeAction(route), route);
+        }
         assertThrows(IllegalArgumentException.class, () -> FleetApp.routeAction("GET /healthz"));
     }
 
+    /**
+     * fleetd #669 Unit A: {@code POST /sessions/{id}/message} is two call shapes behind one route,
+     * mirroring {@code fleet_send}'s MCP-side split into {@link Authz.Action#SEND} and {@link
+     * Authz.Action#ANSWER} ({@code FleetMcp#sendAction}). The route never carries a {@code coordId}
+     * shape — that peer-lead route is MCP-only — so only these two apply here.
+     */
+    @Test
+    void theMessageRouteIsASendWithNoTurnIdAndAnAnswerWithOne() {
+        assertEquals(Authz.Action.SEND, FleetApp.routeAction("POST /sessions/{id}/message", null));
+        assertEquals(Authz.Action.SEND, FleetApp.routeAction("POST /sessions/{id}/message", "  "));
+        assertEquals(Authz.Action.ANSWER, FleetApp.routeAction("POST /sessions/{id}/message", "turn-1"));
+    }
+
+    /**
+     * fleetd #689: {@code answerGatePasses} is the second, conditional gate behind {@code
+     * sendMessage}'s coarse {@link Authz.Action#SEND} check. With the {@code ANSWER} grant denied,
+     * a {@code turnId}-bearing request is refused while a plain one still passes — and the denied
+     * permit is queried only for the {@code turnId} case, never for the plain one, which is what
+     * proves this is a genuinely separate, conditional check rather than the {@code SEND} check
+     * renamed or an unconditional call whose result is ignored. Flipping only the {@code ANSWER}
+     * grant to allowed then flips only the {@code turnId} shape's outcome.
+     */
+    @Test
+    void answerGatePassesOnlyWhenTurnIdAbsentOrAnswerGranted() {
+        List<Authz.Action> queried = new ArrayList<>();
+        Predicate<Authz.Action> denyAnswer = action -> {
+            queried.add(action);
+            return false;
+        };
+
+        assertFalse(FleetApp.answerGatePasses("turn-1", denyAnswer),
+                "ANSWER denied ⇒ the turnId shape is refused");
+        assertEquals(List.of(Authz.Action.ANSWER), queried,
+                "the ANSWER grant, specifically, must be the one consulted");
+
+        queried.clear();
+        assertTrue(FleetApp.answerGatePasses(null, denyAnswer),
+                "no turnId ⇒ the plain shape passes even though ANSWER is denied");
+        assertTrue(FleetApp.answerGatePasses("  ", denyAnswer), "a blank turnId is treated as absent");
+        assertEquals(List.of(), queried, "the plain shape must never consult the permit at all");
+
+        assertTrue(FleetApp.answerGatePasses("turn-1", action -> true),
+                "flipping only the ANSWER grant to allowed flips only the turnId shape's outcome");
+    }
+
     private static Set<String> routesTheServerRegisters() {
         try {
             String source = Files.readString(REST_SOURCE).lines()
@@ -147,6 +203,18 @@ class FleetAppAuthTest {
         }
     }
 
+    /**
+     * {@code permitsFor} is the exact decision {@link FleetApp#allow} makes, passing the real
+     * production classifier rather than a test-supplied one — no terminal is recognised as a
+     * configured lead or collaborator, so a collaborator's SEND is refused through the REST gate.
+     */
+    @Test
+    void aCollaboratorMayNotSendOverRestWithTheRealProductionClassifier() {
+        Principal collaborator = Principal.collaborator("ops", "term_collab", 700);
+        assertFalse(FleetApp.permitsFor(collaborator, Authz.Action.SEND, "term_lead"),
+                "no terminal is recognised as a lead or collaborator yet");
+    }
+
     // --- loopback-trust: the caller is the primary -------------------------------------------
 
     @Test
@@ -192,6 +260,107 @@ class FleetAppAuthTest {
                 "draining an inbox is the primary's collection step");
     }
 
+    /**
+     * fleetd #669 Unit A: the {@code turnId} shape of {@code POST /sessions/{id}/message} maps to
+     * {@link Authz.Action#ANSWER}, not the plain {@link Authz.Action#SEND} the test above drives —
+     * a worker must stay refused on this shape too, exactly as it was refused on the one undivided
+     * action before the split.
+     */
+    @Test
+    void aWorkerMayNotAnswerAnotherSessionsBlockedQuestionOverRest() throws Exception {
+        int port = start(FakeHerdr.WORKER_PID, false, null);
+
+        assertEquals(403, send(port, "POST", "/sessions/term_b/message",
+                "{\"turnId\":\"turn-1\",\"content\":\"hi\"}", null).statusCode(),
+                "resolving another session's blocked question would be a worker escalating too");
+    }
+
+    /**
+     * fleetd #689: a caller refused the coarse {@link Authz.Action#SEND} grant is refused on
+     * {@code SEND} specifically, even on the {@code turnId}-bearing shape that otherwise raises
+     * the check to {@link Authz.Action#ANSWER} — proving {@code turnId} was never read from the
+     * body before the refusal (reading it would have changed which action is named in the 403).
+     * The same caller refused with no body at all gets the identical detail, which could not hold
+     * if the decision depended on anything read from the body. Control: a caller who IS granted
+     * reaches past the gate and the body is used normally.
+     */
+    @Test
+    void aDeniedCallerIsRefusedOnSendEvenWithATurnIdBodyAndNeverReadsTheBody() throws Exception {
+        int workerPort = start(FakeHerdr.WORKER_PID, false, null); // denied: not primary/architect
+
+        HttpResponse<String> withTurnId = send(workerPort, "POST", "/sessions/term_b/message",
+                "{\"turnId\":\"turn-1\",\"content\":\"hi\"}", null);
+        assertEquals(403, withTurnId.statusCode());
+        assertTrue(withTurnId.body().contains("may not SEND"),
+                "the SEND check must be the one that fired, not ANSWER — ANSWER would only be "
+                        + "reachable by having already read turnId out of the body");
+
+        HttpResponse<String> noBody = send(workerPort, "POST", "/sessions/term_b/message", null, null);
+        assertEquals(403, noBody.statusCode());
+        assertTrue(noBody.body().contains("may not SEND"),
+                "refused identically with no body at all — the refusal cannot depend on body content");
+
+        // Control: a primary IS granted SEND, so the same turnId body is read and acted on —
+        // reaching messages.answer, which reports this unknown turnId as a stale one.
+        int primaryPort = start(999_999, false, null);
+        HttpResponse<String> granted = send(primaryPort, "POST", "/sessions/term_b/message",
+                "{\"turnId\":\"turn-1\",\"content\":\"hi\"}", null);
+        assertEquals(409, granted.statusCode());
+        assertTrue(granted.body().contains("stale_turn"), "a granted caller's body IS read and acted on");
+    }
+
+    /**
+     * fleetd #689 (ticket comment 18353): the only place {@code sendMessage}'s call to {@code
+     * answerGatePasses} is observable is the audit trail — {@code allow()} logs an {@code
+     * "allowed"} entry for every granted action except {@code READ}/{@code METRICS}/{@code
+     * TASK_READ}, and {@code ANSWER} is none of those. A granted {@code turnId} request must
+     * therefore log both a {@code SEND} and an {@code ANSWER} entry; a granted plain request must
+     * log {@code SEND} alone. A unit test of the extracted helper pins the helper; this pins the
+     * call site — deleting the {@code answerGatePasses} call from {@code sendMessage} leaves the
+     * helper's own test green but turns this one red.
+     */
+    @Test
+    void aGrantedTurnIdRequestAuditsBothSendAndAnswerButAPlainRequestAuditsSendAlone() throws Exception {
+        int port = start(999_999, false, null); // primary: granted both SEND and ANSWER
+        ObjectMapper mapper = new ObjectMapper();
+
+        try (CapturedLog audit = CapturedLog.at("audit", Level.INFO)) {
+            send(port, "POST", "/sessions/term_b/message",
+                    "{\"turnId\":\"turn-1\",\"content\":\"hi\"}", null);
+
+            List<String> allowed = allowedActions(audit, mapper);
+            assertTrue(allowed.contains("SEND"),
+                    "a turnId request must still clear the coarse SEND grant first");
+            assertTrue(allowed.contains("ANSWER"),
+                    "a turnId request must ALSO clear the ANSWER grant — this is the call site itself");
+        }
+
+        try (CapturedLog audit = CapturedLog.at("audit", Level.INFO)) {
+            send(port, "POST", "/sessions/term_b/message",
+                    "{\"content\":\"hi\",\"timeoutMs\":50}", null);
+
+            List<String> allowed = allowedActions(audit, mapper);
+            assertEquals(List.of("SEND"), allowed,
+                    "a plain request must log SEND and nothing else — ANSWER is conditional on "
+                            + "turnId, not something every request happens to log");
+        }
+    }
+
+    private static List<String> allowedActions(CapturedLog audit, ObjectMapper mapper) {
+        return audit.events().stream()
+                .map(ILoggingEvent::getFormattedMessage)
+                .map(line -> {
+                    try {
+                        return mapper.readTree(line);
+                    } catch (Exception e) {
+                        throw new AssertionError("audit line is not valid JSON: " + line, e);
+                    }
+                })
+                .filter(n -> "allowed".equals(n.path("outcome").asText()))
+                .map(n -> n.path("action").asText())
+                .toList();
+    }
+
     // --- token mode ---------------------------------------------------------------------------
 
     @Test

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

@@ -675,6 +675,26 @@ class FleetAppTest {
         assertEquals(400, postMessage(port, "{}").statusCode());
     }
 
+    /**
+     * fleetd #689: a body that fails to parse is rejected with 400 before {@code turnId} is ever
+     * read from it, so it reaches neither {@code messages.answer} (which needs a {@code turnId})
+     * nor {@code messages.send} — confirmed here for {@code send} by the fake agent's idle status,
+     * which would otherwise make an immediate {@code agent.prompt} delivery observable.
+     */
+    @Test
+    void malformedBodyReturns400AndNeverReachesSendOrAnswer() throws Exception {
+        FakeHerdr herdr = new FakeHerdr().agentStatus("idle"); // idle ⇒ send would deliver right away if reached
+        int port = start(herdr, "http://gx00.gw:8000", Set.of("gx00.gw"));
+
+        HttpResponse<String> res = postMessage(port, "not json at all");
+        assertEquals(400, res.statusCode());
+        JsonNode err = mapper.readTree(res.body());
+        assertEquals("bad_request", err.get("error").asText());
+        assertEquals("body must be JSON", err.get("detail").asText());
+        assertFalse(herdr.called("agent.prompt"),
+                "a malformed body must never reach messages.send's delivery");
+    }
+
     @Test
     void sessionStatusReportsLiveAgentStatus() throws Exception {
         FakeHerdr herdr = new FakeHerdr().agentStatus("blocked");

scripts/config-edit.sh

@@ -213,6 +213,17 @@ map_masked_lines() {
   done < "$file"
 }
 
+# Masks every `scheme://user:pass@host` userinfo on one line of text, replacing just that
+# userinfo with `<redacted>` and leaving the rest of the line untouched, byte for byte. The
+# pattern stops at the first `/`, whitespace, or `@` reached after `://` — a URI's userinfo
+# component cannot contain any of those three characters — so a URI with no userinfo, followed
+# later on the same line by an unrelated `@`, never matches. The `g` flag matters: a line can
+# carry more than one URI. Shared by every caller that prints a line which may hold a
+# credentialed URI, so the bound lives in exactly one place.
+mask_url_userinfo() {
+  printf '%s\n' "$1" | sed -E 's#://[^@/[:space:]]*@#://<redacted>@#g'
+}
+
 redact() {
   local old_file="$1" new_file="$2"
   local line prefix content indent lead key old_line=0 new_line=0 in_hunk=0
@@ -270,7 +281,7 @@ redact() {
         continue
       fi
     fi
-    printf '%s\n' "$line" | sed -E 's#://[^@]*@#://<redacted>@#g'
+    mask_url_userinfo "$line"
   done
   [ "$saved_nocasematch" = 1 ] || shopt -u nocasematch
 }
@@ -580,6 +591,12 @@ install_candidate() {
 
 # -------------------------------------------------------------------------------- the report path
 #
+# Masks basic-auth userinfo (scheme://user:pass@host) in a daemon verdict line before it reaches
+# the terminal.
+mask_verdict_userinfo() {
+  mask_url_userinfo "$1"
+}
+
 # Prints the literal command the operator (or a test) can run to restore the backup by hand — the
 # absolute path to THIS script plus the overrides actually in force, so it works from any cwd.
 restore_command_line() {
@@ -599,8 +616,8 @@ restore_and_confirm() {
   ok "restored from $backup"
   if wait_for_verdict "$LOG" "$mark2" "$WAIT_SECONDS"; then
     case "$VERDICT_KIND" in
-      refused) warn "the RESTORE was also refused by the daemon: $VERDICT_LINE" ;;
+      refused) warn "the RESTORE was also refused by the daemon: $(mask_verdict_userinfo "$VERDICT_LINE")" ;;
-      *)       ok "restore confirmed: $VERDICT_LINE" ;;
+      *)       ok "restore confirmed: $(mask_verdict_userinfo "$VERDICT_LINE")" ;;
     esac
   else
     warn "the restore is on disk, but no confirming verdict line appeared within ${WAIT_SECONDS}s"
@@ -616,7 +633,7 @@ report_outcome() {
 
   say "waiting for the daemon's verdict (up to ${WAIT_SECONDS}s)"
   if wait_for_verdict "$LOG" "$mark" "$WAIT_SECONDS"; then
-    kind="$VERDICT_KIND"; line="$VERDICT_LINE"
+    kind="$VERDICT_KIND"; line="$(mask_verdict_userinfo "$VERDICT_LINE")"
   else
     kind="none"
   fi
@@ -673,7 +690,7 @@ check_mode() {
   local verdict
   verdict="$(last_verdict_line "$LOG")"
   if [ -n "$verdict" ]; then
-    ok "last verdict in log: $verdict"
+    ok "last verdict in log: $(mask_verdict_userinfo "$verdict")"
   else
     warn "no reload verdict line found in $LOG"
   fi

scripts/redeploy-fleetd.sh

@@ -71,20 +71,22 @@ set -euo pipefail
 
 REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 MODULE="$REPO/fleetd"
-JAR="$MODULE/target/fleetd.jar"
+# fleetd #664: the runtime path and Maven's output path are no longer the same file. Maven's
-# fleetd #493: never build into the path a running process holds. The build writes here first
+# shade plugin (finalName=fleetd) always lands a fresh build at target/fleetd.jar — that is
-# (Maven's shade plugin has finalName=fleetd, so `clean install` still lands its output at
+# Maven's own output directory and this script does not change it — but the daemon is launched
-# target/fleetd.jar — that part is unchanged and out of this script's control), but this script
+# from $JAR instead, outside target/ entirely. That split is the whole fix: neither `mvn install`
-# now moves it out to JAR_STAGED immediately, and only swaps it back to JAR (a plain `mv`, so a
+# nor `mvn clean` can ever reach the file a running daemon holds open, because that file no
-# rename, never a byte-by-byte overwrite) after the OLD daemon has been confirmed exited. See
+# longer lives under target/ at all. See swap_if_built/swap_staged_jar below for the one `mv`
-# stage_built_jar/swap_staged_jar below.
+# that moves a build from one path to the other, and only after the old daemon is confirmed gone.
-JAR_STAGED="$MODULE/target/fleetd-new.jar"
+BUILD_JAR="$MODULE/target/fleetd.jar"
+JAR="$MODULE/run/fleetd.jar"
 OUT="$MODULE/fleetd.out"
-# Matches BOTH the absolute form and the relative `java -jar target/fleetd.jar` a hand-start
+# Matches a fleetd daemon's command line wherever its jar sits — absolute or relative, under
-# produces from inside fleetd/. Anchoring on the absolute path alone was a real bug: the daemon
+# run/, under target/, or anywhere else a build or a hand-start might point it. Detecting a
-# restarted correctly and the script still reported "no process appeared", because it launched with
+# daemon this script did not start, including one running from a jar outside $JAR's own
-# a relative path and then looked for an absolute one.
+# directory, is this pattern's whole job; running_pid()'s `comm = java` allowlist below is what
-PATTERN='target/fleetd.jar'
+# keeps that breadth from counting a shell that merely types the pattern as literal text.
+PATTERN='fleetd.jar'
 HEALTH='http://127.0.0.1:8765/healthz'
 STOP_WAIT=30      # seconds to wait for a clean exit before reporting failure
 HEALTH_WAIT=60    # seconds to wait for /healthz to answer after start — fleetd #603: also the pid-
@@ -169,10 +171,10 @@ hash256() {
   fi
 }
 
-# Reports the hash of $JAR by default, or of whatever path is passed — used to report the STAGED
+# Reports the hash of $JAR by default, or of whatever path is passed — used to report the BUILT
-# jar right after a build (before it has been swapped in) without ever changing what a bare
+# jar at $BUILD_JAR (before it has been swapped in) without ever changing what a bare `jar_id`
-# `jar_id` (no args) means: the live path, $JAR. --check and the final "pid ..., jar ..." line
+# (no args) means: the live path, $JAR. --check and the final "pid ..., jar ..." line both call
-# both call it with no args on purpose, so neither can ever be fooled by a leftover staged file.
+# it with no args on purpose, so neither can ever be fooled by a leftover build output.
 # fleetd #550 — THREE distinct answers now, not two: `[ -f "$f" ]` already separates "the jar is
 # not there" (-> "absent") from "the jar is there"; for the second case, hash256 itself separates
 # "hashed it" (a 12-char hex string) from "could not hash it" (-> "unhashable", when no hasher is
@@ -180,15 +182,35 @@ hash256() {
 # was the whole defect this ticket fixes.
 jar_id() { local f="${1:-$JAR}"; [ -f "$f" ] && hash256 "$f" || echo "absent"; }
 
+# fleetd #664 — under the old layout $JAR and the build output were the same file, so "jar on
+# disk" was one fact. Now they are two: $BUILD_JAR (target/fleetd.jar, whatever Maven last wrote,
+# by this script or by a bare `mvn install` run by hand) and $JAR (run/fleetd.jar, whatever the
+# daemon actually has open). Printing one label for both was the trap this ticket exists to close
+# — during the incident it would have shown the NEW jar's hash while the JVM ran the OLD one.
+# Pure (reads jar_id/date, never mutates), so a test can call it directly without reaching the
+# main flow — the same shape swap_if_built/drain_gate_refusal already use.
+report_jar_state() {
+  local built_hash running_hash built_mtime running_mtime
+  built_hash="$(jar_id "$BUILD_JAR")"
+  running_hash="$(jar_id "$JAR")"
+  built_mtime="$([ -f "$BUILD_JAR" ] && date -r "$BUILD_JAR" '+%Y-%m-%d %H:%M:%S' || echo 'none')"
+  running_mtime="$([ -f "$JAR" ] && date -r "$JAR" '+%Y-%m-%d %H:%M:%S' || echo 'none')"
+  ok "built jar   (target/fleetd.jar): $built_hash  ($built_mtime)"
+  ok "running jar (run/fleetd.jar):    $running_hash  ($running_mtime)"
+  if [ "$built_hash" != "absent" ] && [ "$running_hash" != "absent" ] && [ "$built_hash" != "$running_hash" ]; then
+    warn "built jar and running jar differ — target/fleetd.jar was rebuilt since the running daemon last started and is not yet live"
+  fi
+}
+
 # fleetd #593 — `pgrep -f "$PATTERN"` matches ANY process whose full command line CONTAINS the
 # pattern text, and that is not the same thing as "is the daemon". A shell that merely embeds the
 # pattern as literal text — a human typing this exact investigation by hand, an ssh-shaped
 # `sh -c '...; ...'`, a pipeline, or any other non-exec'ing shell that never replaced itself with
 # the pattern-holding command — still shows up in that match, and it is the INSTRUMENT, not the
-# daemon. Measured live on this Mac: `sh -c 'echo "target/fleetd.jar" >/dev/null; sleep 30' &`
+# daemon. Measured live on this Mac: `sh -c 'echo "run/fleetd.jar" >/dev/null; sleep 30' &`
 # leaves a real `sh` process alive (it forks for the `sleep`, it does not exec into it) whose own
-# `ps -o args` is `sh -c echo "target/fleetd.jar" >/dev/null; sleep 30` — `pgrep -f "$PATTERN"`
+# `ps -o args` is `sh -c echo "run/fleetd.jar" >/dev/null; sleep 30` — `pgrep -f "$PATTERN"`
-# matches that line right alongside the real `java -jar target/fleetd.jar` process. `pgrep -c`
+# matches that line right alongside the real `java -jar run/fleetd.jar` process. `pgrep -c`
 # (an in-one-call count) does not exist on BSD/macOS at all, so this cannot be fixed by switching
 # pgrep flags — it has to filter what pgrep already found, after the fact, in a way that still
 # runs on BSD.
@@ -210,7 +232,7 @@ jar_id() { local f="${1:-$JAR}"; [ -f "$f" ] && hash256 "$f" || echo "absent"; }
 # launched as `java -jar ...` — a native image, a renamed launcher — `running_pid()` silently
 # returns nothing and `assert_single_daemon` stops noticing a second daemon at all. For a guard,
 # that false-negative direction is the worse one to be wrong in. This is not a new assumption,
-# though: `PATTERN='target/fleetd.jar'` two lines up already assumes the daemon is a jar, which
+# though: `PATTERN='fleetd.jar'` two lines up already assumes the daemon is a jar, which
 # is only ever run by `java`. If that launch method changes, `PATTERN` stops matching anything
 # before this allowlist would ever get the chance to be wrong — the allowlist rides on the same
 # assumption that is already load-bearing, it does not add a new one. Whoever changes the launch
@@ -229,16 +251,12 @@ running_pid() {
   printf '%s' "$out"
 }
 
-# fleetd #493 — three small, independently testable pieces of "never build into the path a
+# fleetd #493/#664 — the independently testable pieces of "never build into the path a running
-# running process holds":
+# process holds":
 #
-#   stage_built_jar   moves the jar Maven just produced OUT of the live path and onto the staging
+#   require_no_build_jar   the --no-build path never builds anything: it must find a jar already
-#                      path, immediately after a successful build. Dies (leaving the OLD daemon
+#                      sitting at the live path ($JAR, under run/) from an earlier successful run,
-#                      untouched — this runs before the stop step) if Maven reported success but
+#                      and die with the same truthful message this script has always used if not.
-#                      left no jar behind, or if the move itself fails.
-#   require_no_build_jar   the --no-build path never builds or stages anything: it must find a
-#                      jar already sitting at the live path from an earlier successful run, and
-#                      die with the same truthful message this script has always used if not.
 #   wait_for_daemon_exit   polls running_pid() for up to $1 seconds and reports whether the OLD
 #                      daemon actually exited — extracted to its own function so the main flow
 #                      can be relied on to call swap_staged_jar only AFTER this returns success,
@@ -248,14 +266,10 @@ running_pid() {
 #                      so this is never a write into a path a running process holds — by the time
 #                      it runs, nothing holds that path anymore. If it fails, the caller must not
 #                      start a new daemon: die() below already refuses that by exiting the script.
-stage_built_jar() {
+#                      fleetd #664: the "staged" jar swap_if_built passes in is now $BUILD_JAR
-  [ -f "$JAR" ] || die "build succeeded but produced no jar at $JAR — cannot stage it for restart.
+#                      itself (target/fleetd.jar, Maven's own output) — a build no longer needs to
-      The running daemon was NOT touched."
+#                      be moved off the live path right after compiling, because target/ was never
-  mv -f "$JAR" "$JAR_STAGED" \
+#                      the live path to begin with.
-    || die "could not move the freshly built jar from $JAR to the staging path $JAR_STAGED.
-        The running daemon was NOT touched."
-}
-
 require_no_build_jar() {
   [ -f "$JAR" ] || die "no jar at $JAR — run without --no-build"
 }
@@ -282,8 +296,7 @@ swap_staged_jar() {
 #
 # The defect: the swap step used to be guarded inline by `if [ "$DO_BUILD" = 1 ]` in the main flow.
 # Changing that to `if false` left the suite green and the swap never ran, so a redeploy reported
-# every step succeeding while the daemon started on no jar at all (stage_built_jar has already moved
+# every step succeeding while the daemon started on no jar at all or on a stale one.
-# the freshly built one to $JAR_STAGED by then) or on a stale one.
 # test_swap_ordered_after_wait_and_before_start could not catch it: it reads this script's own text
 # and compares line positions, and a same-line edit moves no line.
 #
@@ -312,7 +325,12 @@ swap_if_built() {
   local do_build="$1"
   should_swap "$do_build" || return 0
   say "swap"
-  swap_staged_jar "$JAR_STAGED" "$JAR"
+  # fleetd #664: $JAR now lives under run/, a directory target/ never created. mkdir -p here,
+  # not inside swap_staged_jar itself — that function's own contract is tested on a missing
+  # parent directory (a failing mv), and widening it to auto-create one would change what that
+  # test proves.
+  mkdir -p "$(dirname "$JAR")"
+  swap_staged_jar "$BUILD_JAR" "$JAR"
   ok "jar in place: $(jar_id)"
 }
 
@@ -610,6 +628,45 @@ check_log_path_matches_plist() {
   ok "log path check: script and plist agree ($resolved_out)"
 }
 
+# Reads the launchd plist's ProgramArguments for the argument that follows "-jar", resolves it
+# alongside $jar_path, and dies when the two differ. Call it only when the agent is loaded; it
+# never touches launchd or the daemon itself.
+check_jar_path_matches_plist() {
+  local jar_path="$1" plist_path="$2"
+  local plist_args plist_jar resolved_jar resolved_plist_jar
+  if ! plist_args="$(/usr/libexec/PlistBuddy -c 'Print :ProgramArguments' "$plist_path" 2>/dev/null)"; then
+    die "launchd agent is loaded but PlistBuddy could not read ProgramArguments from
+        $plist_path
+      — cannot verify which jar the supervised daemon launches. Fix the plist before
+      redeploying supervised."
+  fi
+  plist_jar="$(printf '%s\n' "$plist_args" | awk '
+      { gsub(/^[ \t]+|[ \t]+$/, "") }
+      prev == "-jar" { print; exit }
+      { prev = $0 }
+    ')"
+  if [ -z "$plist_jar" ]; then
+    die "launchd agent is loaded but its ProgramArguments at
+        $plist_path
+      do not contain a '-jar <path>' pair — cannot verify which jar the supervised daemon
+      launches. Fix the plist before redeploying supervised."
+  fi
+  resolved_jar="$(cd "$(dirname "$jar_path")" 2>/dev/null && pwd -P)/$(basename "$jar_path")" || true
+  resolved_plist_jar="$(cd "$(dirname "$plist_jar")" 2>/dev/null && pwd -P)/$(basename "$plist_jar")" || true
+  if [ -z "$resolved_jar" ] || [ -z "$resolved_plist_jar" ] || [ "$resolved_jar" != "$resolved_plist_jar" ]; then
+    die "jar path mismatch — this script deploys to
+        $jar_path  (resolved: ${resolved_jar:-<directory does not exist>})
+      but the loaded plist's ProgramArguments names
+        $plist_jar  (resolved: ${resolved_plist_jar:-<directory does not exist>})
+      The swap renames the built jar into place, so the old path stops existing after a redeploy;
+      a launchd-initiated start from this plist (a reboot, or KeepAlive after a crash) would then
+      run java against a missing file. Reinstall the plist at
+        $plist_path
+      so its ProgramArguments names $jar_path before redeploying supervised."
+  fi
+  ok "jar path check: script and plist agree ($resolved_jar)"
+}
+
 # fleetd #552: the post-restart fresh-log capture, pulled out of the main flow so it is testable by
 # sourcing (the same reason systemd_installed/systemd_loaded above guard their OWN mktemp inline
 # instead of leaving it bare) even though its only caller sits below the SOURCED guard. By the time
@@ -826,16 +883,25 @@ report_shutdown_drain() {
 #   --no-build, staged jar present -> ALSO "nothing changed", deliberately: --no-build itself builds
 #     and stages nothing (see require_no_build_jar above), so a staged jar found here is a leftover
 #     from an earlier, unrelated run. THIS run truly changed nothing, and the next DO_BUILD=1 run
-#     wipes that leftover before it builds (`rm -f "$JAR_STAGED"` in the build section above) — so
+#     wipes that leftover for free — `mvn clean` deletes all of target/, $BUILD_JAR included,
-#     there is nothing here for the operator to lose track of.
+#     before the build even starts — so there is nothing here for the operator to lose track of.
 #   --no-build, staged jar absent  -> "nothing changed"
 drain_gate_refusal() {
   local do_build="$1" staged_path="$2"
   if [ "$do_build" = 1 ] && [ -f "$staged_path" ]; then
+    # fleetd #664: under the old layout a build emptied the live path ($JAR) immediately, so
+    # --no-build's own check ("no jar at $JAR") was the thing that refused a rerun here. That is
+    # no longer true: a build never touches $JAR at all now, so $JAR still holds whatever was
+    # already running before this gate fired (reaching this message at all requires OLD_PID to
+    # have been set, which means a daemon was running from $JAR already) — a --no-build rerun
+    # would NOT refuse, it would just restart that same old jar and silently throw away the one
+    # sitting at %s.
     printf 'aborted — the running daemon was NOT touched, but the freshly built jar is sitting at
-        %s, not yet swapped into %s. Rerun WITHOUT --no-build to finish the restart —
+        %s, not yet swapped into %s. Rerun WITHOUT --no-build to finish the restart — a
-        the freshly built jar is no longer at the live path that --no-build requires — or
+        --no-build rerun would NOT refuse here: %s already exists from before this run, so it
-        remove %s by hand if you want to discard this build.' "$staged_path" "$JAR" "$staged_path"
+        would restart the daemon on that OLD jar and silently discard the one you just built — or
+        remove %s by hand if you want to discard this build instead.' \
+      "$staged_path" "$JAR" "$JAR" "$staged_path"
   else
     printf 'aborted — nothing changed'
   fi
@@ -934,6 +1000,7 @@ report_supervisor_state() {
       SUPERVISED=1
       ok "launchd agent loaded ($LAUNCHD_LABEL) — launchd supervises this daemon"
       check_log_path_matches_plist "$OUT" "$LAUNCHD_PLIST"
+      check_jar_path_matches_plist "$JAR" "$LAUNCHD_PLIST"
       ;;
     systemd)
       SUPERVISED=1
@@ -1170,7 +1237,7 @@ if [ -n "$OLD_PID" ]; then
 else
   warn "no daemon running — this will be a cold start"
 fi
-ok "jar on disk: $(jar_id)  ($([ -f "$JAR" ] && date -r "$JAR" '+%Y-%m-%d %H:%M:%S' || echo 'none'))"
+report_jar_state
 ok "HEAD: $(git -C "$REPO" log --oneline -1)"
 
 # CB-594 / fleetd #492: supervision state. Installed and loaded are different facts — a
@@ -1252,9 +1319,6 @@ stop_if_check_only "$CHECK_ONLY"
 
 if [ "$DO_BUILD" = 1 ]; then
   say "build"
-  # fleetd #493: wipe a leftover staged jar from a previous failed/interrupted run BEFORE doing
-  # anything else, so that run's leftovers can never be mistaken for this run's output.
-  rm -f "$JAR_STAGED"
   BUILD_LOG="$(mktemp -t fleetd-build.XXXXXX)"
   echo "   log: $BUILD_LOG"
   if ! mvn -f "$MODULE/pom.xml" clean install > "$BUILD_LOG" 2>&1; then
@@ -1264,16 +1328,16 @@ if [ "$DO_BUILD" = 1 ]; then
   fi
   grep -E '^\[INFO\] Tests run:.*Failures' "$BUILD_LOG" | tail -1 | sed 's/^\[INFO\] /   /' || true
   ok "BUILD SUCCESS"
-  # fleetd #493: move the freshly built jar off the live path immediately — the running (OLD)
+  # fleetd #664: nothing to stage — $BUILD_JAR (target/fleetd.jar) is Maven's own output path and
-  # daemon, if any, is still up at this point (build always runs before stop). From here until the
+  # was never the live path, so the running (OLD) daemon, if any, was never at risk from this build
-  # swap step below (after the OLD daemon is confirmed gone), $JAR_STAGED is the only artefact this
+  # at all. From here until the swap step below (after the OLD daemon is confirmed gone),
-  # script treats as "the new jar" — $JAR itself is not touched again until the swap.
+  # $BUILD_JAR is the artefact this script treats as "the new jar" — $JAR itself is not touched
-  stage_built_jar
+  # again until the swap.
-  ok "jar now: $(jar_id "$JAR_STAGED")"
+  ok "jar now: $(jar_id "$BUILD_JAR")"
 else
   say "build skipped (--no-build)"
-  # fleetd #493: --no-build never builds or stages anything — it restarts whatever jar is already
+  # fleetd #493: --no-build never builds anything — it restarts whatever jar is already sitting at
-  # sitting at the live path from an earlier successful run. Same check, same message as before.
+  # the live path from an earlier successful run. Same check, same message as before.
   require_no_build_jar
 fi
 
@@ -1282,8 +1346,8 @@ fi
 # fleetd #555: run_drain_gate above is drain_gate_required + the prompt + drain_confirmed, called
 # unconditionally — it returns immediately when the gate is not required, and composes/dies through
 # refuse_drain_gate itself when the reply does not confirm. See #493/#517/#528 for why "nothing
-# changed" would be a lie once a build has staged a jar.
+# changed" would be a lie once a build has produced a jar not yet swapped in.
-run_drain_gate "$OLD_PID" "$ASSUME_YES" "$DO_BUILD" "$JAR_STAGED"
+run_drain_gate "$OLD_PID" "$ASSUME_YES" "$DO_BUILD" "$BUILD_JAR"
 
 # ------------------------------------------------------------------ stop
 #
@@ -1334,21 +1398,22 @@ fi
 
 # ------------------------------------------------------------------ swap
 #
-# fleetd #493: every branch above has now either confirmed the OLD daemon actually exited
+# fleetd #493/#664: every branch above has now either confirmed the OLD daemon actually exited
-# (wait_for_daemon_exit, above) or established there was never one running to begin with. Only
+# (wait_for_daemon_exit, above) or established there was never one running to begin with. Only NOW
-# NOW is it safe to put the freshly built jar at the path the NEXT `java -jar` (direct, or via
+# is it safe to put the freshly built jar at the path the NEXT `java -jar` (direct, or via
-# launchd/systemd's ExecStart) will read from — this mv is the one and only write to $JAR anywhere
+# launchd/systemd's ExecStart) will read from — a rename from $BUILD_JAR (target/) to $JAR (run/),
+# both under $MODULE and so on one filesystem. This mv is the one and only write to $JAR anywhere
 # in this script's mutating flow. If it fails, do not start: die() below exits before "start" runs.
 swap_if_built "$DO_BUILD"
 
 # ------------------------------------------------------------------ start
 # Unsupervised: login shell (zsh -l) is what puts the secrets on the daemon's environment, and cwd
-# must be fleetd/ because the daemon resolves fleetd.yaml, logs/ and target/ relative to it.
+# must be fleetd/ because the daemon resolves fleetd.yaml, logs/ and run/ relative to it.
 # Supervised (launchd): launchd does both — deploy/dev.ltms.fleetd.plist points ProgramArguments at
 # scripts/fleetd-launchd-wrapper.sh (CB-594), which is what execs the login shell in launchd's
 # place, and WorkingDirectory in the plist already pins fleetd/.
 # Supervised (systemd --user): the unit does both too — measured on the second host, ExecStart is
-# `/bin/zsh -lc "exec java -jar target/fleetd.jar fleetd.yaml"` (a login shell, same reason as
+# `/bin/zsh -lc "exec java -jar run/fleetd.jar fleetd.yaml"` (a login shell, same reason as
 # above) and WorkingDirectory is already pinned to fleetd/.
 
 say "start"

scripts/test-config-edit.sh

@@ -233,6 +233,66 @@ test_redaction_holds() {
   assert_contains "weight" "$RUN_OUTPUT" "a diff must have been demonstrably printed at all"
 }
 
+# redact()'s key-name filter only inspects the KEY, so a diff line whose key does not match
+# TOKEN|SECRET|PASSWORD|PASSWD|PASSPHRASE|CREDENTIAL|URI|_KEY still reaches the final userinfo
+# sed even when its VALUE holds a credentialed URI. "note" is not a sensitive key name, so this
+# line must fall all the way through to that sed, not the earlier whole-value branch. The
+# trailing prose on both sides of the userinfo is a positive control: it proves the line reached
+# the userinfo sed (which touches only the userinfo) rather than the earlier branch (which would
+# have replaced the whole value with a bare "<redacted>" and dropped the prose).
+test_diff_line_userinfo_is_masked_with_positive_control() {
+  local dir
+  dir="$(new_fixture)"
+
+  start_run "$dir" 5 --set '.profiles.sonnet.note=see amqp://alice:wonderland@rabbit.local:5672/vhost for details'
+  sleep 1
+  printf 'config reloaded\n' >> "$dir/fleetd.out"
+  collect_run "$dir"
+
+  assert_equals 0 "$RUN_RC" "diff-userinfo-case reload exit code"
+  assert_not_contains "alice:wonderland" "$RUN_OUTPUT" "the userinfo must never reach the output"
+  assert_contains "amqp://<redacted>@rabbit.local:5672/vhost" "$RUN_OUTPUT" \
+    "the userinfo must be MASKED, not deleted — the rest of the value must survive"
+  assert_contains "note:" "$RUN_OUTPUT" "the key name must still reach the output"
+  assert_contains "see " "$RUN_OUTPUT" "prose BEFORE the userinfo must still reach the output"
+  assert_contains "for details" "$RUN_OUTPUT" "prose AFTER the userinfo must still reach the output"
+}
+
+# A diff line can hold a URL with no userinfo, followed later on the same line by an unrelated @
+# (free text in a string value, for example an email address). The line must pass through the
+# userinfo sed byte for byte: the match must stop at the end of the URL and must not treat the
+# later @ as a second userinfo delimiter.
+test_diff_line_uri_without_userinfo_survives_a_later_at_sign() {
+  local dir
+  dir="$(new_fixture)"
+
+  start_run "$dir" 5 --set '.profiles.sonnet.note2=see https://docs.local/guide and mail ops@example.com'
+  sleep 1
+  printf 'config reloaded\n' >> "$dir/fleetd.out"
+  collect_run "$dir"
+
+  assert_equals 0 "$RUN_RC" "diff-no-userinfo-with-later-at-sign reload exit code"
+  assert_contains "note2: see https://docs.local/guide and mail ops@example.com" "$RUN_OUTPUT" \
+    "a URL with no userinfo plus a later @ on the same line must pass through byte for byte"
+}
+
+# Two credentialed URIs on one diff line must both be masked — the g flag matters.
+test_diff_line_masks_multiple_userinfo_with_g_flag() {
+  local dir
+  dir="$(new_fixture)"
+
+  start_run "$dir" 5 --set '.profiles.sonnet.note3=amqp://u1:p1@host1/vhost1 and amqp://u2:p2@host2/vhost2'
+  sleep 1
+  printf 'config reloaded\n' >> "$dir/fleetd.out"
+  collect_run "$dir"
+
+  assert_equals 0 "$RUN_RC" "diff-two-userinfo-on-one-line reload exit code"
+  assert_not_contains "u1:p1" "$RUN_OUTPUT" "the first userinfo must never reach the output"
+  assert_not_contains "u2:p2" "$RUN_OUTPUT" "the second userinfo must never reach the output"
+  assert_contains "amqp://<redacted>@host1/vhost1" "$RUN_OUTPUT" "the first URI must be masked"
+  assert_contains "amqp://<redacted>@host2/vhost2" "$RUN_OUTPUT" "the second URI must be masked"
+}
+
 # ------------------------------------------------------- acceptance criterion 9: forgotten value
 # `--set .a.b=` is a plausible typo (the value simply forgotten), and it must be refused outright
 # rather than silently nulling the field — a null numeric field falls back to its default, which
@@ -629,6 +689,65 @@ test_refusal_shape_from_parse_failure_wording_is_recognised() {
   assert_equals 4 "$RUN_RC" "the parse-failure refusal shape must also exit 4, not be read as silence"
 }
 
+# A verdict line carrying a credentialed URI has its userinfo masked, with a positive control
+# proving the rest of the line still reaches the output unchanged.
+test_verdict_userinfo_is_masked_with_positive_control() {
+  local dir
+  dir="$(new_fixture)"
+
+  start_run "$dir" 5 --set '.profiles.sonnet.weight=4'
+  sleep 1
+  printf 'config reload from %s refused, keeping the running config: refusing to start: malformed pattern — profiles.local.errorPattern ("amqp://user:hunter2@host/vhost"): Unclosed character class near index 8\n' \
+    "$dir/fleetd.yaml" >> "$dir/fleetd.out"
+  collect_run "$dir"
+
+  assert_equals 4 "$RUN_RC" "refusal-with-userinfo exit code"
+  assert_not_contains "user:hunter2" "$RUN_OUTPUT" "the userinfo must never reach the output"
+  assert_contains "amqp://<redacted>@host/vhost" "$RUN_OUTPUT" \
+    "the userinfo must be MASKED, not deleted — the rest of the quoted value must survive"
+  # Positive control: the diagnostic prose on both sides of the userinfo must still reach the
+  # output. Without this, a mutant that drops the whole verdict line would pass identically.
+  assert_contains "malformed pattern" "$RUN_OUTPUT" "prose BEFORE the userinfo must still reach the output"
+  assert_contains "Unclosed character class near index 8" "$RUN_OUTPUT" \
+    "prose AFTER the userinfo must still reach the output"
+}
+
+# An ordinary refusal line quotes the offending pattern, not a credential, and must survive byte
+# for byte: the rewrite is scoped to userinfo only, and the quoted pattern is the detail an
+# operator needs to fix the refusal.
+test_ordinary_refusal_line_passes_through_unchanged() {
+  local dir real_line
+  dir="$(new_fixture)"
+  real_line="config reload from $dir/fleetd.yaml refused, keeping the running config: refusing to start: malformed pattern — profiles.local.errorPattern (\"[unclosed\"): Unclosed character class near index 8"
+
+  start_run "$dir" 5 --set '.profiles.sonnet.weight=4'
+  sleep 1
+  printf '%s\n' "$real_line" >> "$dir/fleetd.out"
+  collect_run "$dir"
+
+  assert_equals 4 "$RUN_RC" "ordinary refusal exit code"
+  assert_contains "$real_line" "$RUN_OUTPUT" \
+    "an ordinary refusal with no userinfo must pass through byte for byte, unchanged"
+}
+
+# A verdict line can hold a URI with NO userinfo and a later, unrelated @ further on in the same
+# line (an email address in diagnostic prose, for example). The rewrite must stop at the end of
+# the URI and must not treat the later @ as a second userinfo delimiter.
+test_uri_without_userinfo_survives_a_later_at_sign() {
+  local dir real_line
+  dir="$(new_fixture)"
+  real_line="config reload from $dir/fleetd.yaml refused, keeping the running config: broker.uri amqp://broker.local/vhost unreachable, contact ops@example.com"
+
+  start_run "$dir" 5 --set '.profiles.sonnet.weight=4'
+  sleep 1
+  printf '%s\n' "$real_line" >> "$dir/fleetd.out"
+  collect_run "$dir"
+
+  assert_equals 4 "$RUN_RC" "no-userinfo-with-later-at-sign exit code"
+  assert_contains "$real_line" "$RUN_OUTPUT" \
+    "a URI with no userinfo plus a later @ in the same line must pass through byte for byte"
+}
+
 # --set runs yq over the whole candidate. It warns when that changes more lines than the requested
 # pairs, but a simple file with only the intended changed line must stay quiet.
 new_fixture_reformat_sensitive() {
@@ -694,6 +813,12 @@ echo "== acceptance criterion 6: the marker works =="
 test_marker_skips_lines_before_it
 echo "== acceptance criterion 7 (+13: redaction is proven to have run) =="
 test_redaction_holds
+echo "== fleetd #692: a diff line's userinfo is masked, rest of the value survives =="
+test_diff_line_userinfo_is_masked_with_positive_control
+echo "== fleetd #692: a diff line's URI with no userinfo survives a later @ in the line =="
+test_diff_line_uri_without_userinfo_survives_a_later_at_sign
+echo "== fleetd #692: two userinfo URIs on one diff line are both masked =="
+test_diff_line_masks_multiple_userinfo_with_g_flag
 echo "== acceptance criterion 9: a forgotten value refuses and installs nothing =="
 test_forgotten_value_refuses_and_installs_nothing
 echo "== acceptance criterion 10: an explicit clear writes a bare null =="
@@ -720,6 +845,12 @@ echo "== extra: --check is read-only and always exits 0 =="
 test_check_is_read_only_and_exits_zero
 echo "== extra: the parse-failure refusal shape is also recognised =="
 test_refusal_shape_from_parse_failure_wording_is_recognised
+echo "== verdict-redaction criteria 2+3: verdict userinfo is masked, rest of line survives =="
+test_verdict_userinfo_is_masked_with_positive_control
+echo "== verdict-redaction criterion 4: an ordinary refusal passes through unchanged =="
+test_ordinary_refusal_line_passes_through_unchanged
+echo "== fleetd #638: a URI with no userinfo survives a later @ in the same line =="
+test_uri_without_userinfo_survives_a_later_at_sign
 echo "== acceptance criterion 17: --set warns about yq formatting churn =="
 test_set_warns_when_yq_reformats_extra_lines
 echo "== acceptance criterion 18: --set stays quiet without formatting churn =="

scripts/test-redeploy-fleetd.sh

@@ -447,7 +447,7 @@ test_assert_single_daemon_rejects_two_pids() {
 test_running_pid_excludes_self_matching_wrapper_shell() {
   local before after wrapper_pid
   before="$(running_pid)"
-  sh -c 'echo "target/fleetd.jar" >/dev/null; sleep 20' &
+  sh -c 'echo "run/fleetd.jar" >/dev/null; sleep 20' &
   wrapper_pid=$!
   sleep 0.3
   after="$(running_pid)"
@@ -475,11 +475,26 @@ test_running_pid_excludes_self_matching_wrapper_shell() {
 # `comm` from the actually-executed binary's own path, not from `exec -a`'s argv[0] override (BSD
 # ties `comm` to argv[0], which is what makes this technique work here) — so on Linux this
 # specific fixture might report `comm=sh`, not `comm=java`, even though the REAL daemon (a literal
-# `java -jar target/fleetd.jar` process, never fabricated) is unaffected either way. I could not
+# `java -jar run/fleetd.jar` process, never fabricated) is unaffected either way. I could not
 # verify this fixture's behavior on Linux, so test_running_pid_counts_a_pid_whose_comm_is_java
 # below backstops the same claim (the allowlist admits a pid whose comm is `java`) with a stubbed
 # `ps`, which is identical bash on every platform and carries no such platform question.
 test_running_pid_finds_a_real_java_named_second_process() {
+  local before after standin_pid
+  before="$(running_pid)"
+  ( exec -a java sh -c 'echo "run/fleetd.jar" >/dev/null; sleep 20' ) &
+  standin_pid=$!
+  sleep 0.3
+  after="$(running_pid)"
+  kill "$standin_pid" 2>/dev/null || true
+  wait "$standin_pid" 2>/dev/null || true
+  printf '%s\n' "$after" | grep -qxF "$standin_pid" \
+    || fail "running_pid() did not find a real second process (pid $standin_pid, comm forced to 'java' via exec -a) whose own argv holds the pattern: before=[$before] after=[$after]"
+}
+
+# PATTERN matches a fleetd jar in either build layout, not only the run/ one: a process whose
+# argv names a jar under target/ must be found too, the same way the run/ case above is.
+test_running_pid_finds_a_real_java_named_process_from_target_dir() {
   local before after standin_pid
   before="$(running_pid)"
   ( exec -a java sh -c 'echo "target/fleetd.jar" >/dev/null; sleep 20' ) &
@@ -489,7 +504,24 @@ test_running_pid_finds_a_real_java_named_second_process() {
   kill "$standin_pid" 2>/dev/null || true
   wait "$standin_pid" 2>/dev/null || true
   printf '%s\n' "$after" | grep -qxF "$standin_pid" \
-    || fail "running_pid() did not find a real second process (pid $standin_pid, comm forced to 'java' via exec -a) whose own argv holds the pattern: before=[$before] after=[$after]"
+    || fail "running_pid() did not find a real second process (pid $standin_pid, comm forced to 'java' via exec -a) naming a jar under target/: before=[$before] after=[$after]"
+}
+
+# Broadening PATTERN to match both build layouts must not also broaden it into matching a
+# non-exec'ing shell that merely holds the target/ text as a literal argument, the same
+# self-matching shape test_running_pid_excludes_self_matching_wrapper_shell above excludes for
+# the run/ text.
+test_running_pid_excludes_self_matching_wrapper_shell_naming_target_dir() {
+  local before after wrapper_pid
+  before="$(running_pid)"
+  sh -c 'echo "target/fleetd.jar" >/dev/null; sleep 20' &
+  wrapper_pid=$!
+  sleep 0.3
+  after="$(running_pid)"
+  kill "$wrapper_pid" 2>/dev/null || true
+  wait "$wrapper_pid" 2>/dev/null || true
+  [ "$after" = "$before" ] \
+    || fail "running_pid() counted a self-matching wrapper shell (pid $wrapper_pid, holding 'target/fleetd.jar' as literal text in its own argv, not the daemon): before=[$before] after=[$after]"
 }
 
 # fleetd #593 CORRECTION 1, hole 2 — the round-1 filter denied known shell names (sh/bash/zsh/
@@ -557,29 +589,30 @@ test_die_message_does_not_recommend_bare_pgrep_as_remediation() {
     || fail "assert_single_daemon's die message does not say in words that a pattern can match the caller (fleetd #593)"
 }
 
-# fleetd #511 — jar_id()'s no-argument default was unpinned by any test: nothing proved it reports
+# fleetd #511/#664 — jar_id()'s no-argument default was unpinned by any test: nothing proved it
-# $JAR (the live path) rather than $JAR_STAGED. Both halves matter, so this pins both: the bare call
+# reports $JAR (the live path) rather than whatever explicit path a caller passes it (e.g.
-# must hash the live jar, and an explicit path argument must hash THAT file, not fall back to $JAR.
+# $BUILD_JAR). Both halves matter, so this pins both: the bare call must hash the live jar, and an
-# Two files with different content, so a default pointed at the wrong one reports the wrong hash
+# explicit path argument must hash THAT file, not fall back to $JAR. Two files with different
-# rather than accidentally matching.
+# content, so a default pointed at the wrong one reports the wrong hash rather than accidentally
+# matching.
 test_jar_id_defaults_to_live_and_reports_explicit_path() {
-  local dir saved_jar="$JAR" saved_staged="$JAR_STAGED"
+  local dir saved_jar="$JAR"
-  local live_hash staged_hash default_result explicit_result
+  local live_hash other_hash default_result explicit_result other_path
   dir="$TMP/jar-id"; mkdir -p "$dir"
-  JAR="$dir/fleetd.jar"; JAR_STAGED="$dir/fleetd-new.jar"
+  JAR="$dir/fleetd.jar"; other_path="$dir/other.jar"
   printf 'live jar bytes' > "$JAR"
-  printf 'staged jar bytes, not the same content' > "$JAR_STAGED"
+  printf 'other jar bytes, not the same content' > "$other_path"
   # fleetd #550: this reference hash must be computed the same portable way jar_id() itself now
   # computes one — a bare, unguarded call to the macOS-only hasher here was exactly the item-2
   # defect, dying with "command not found" on any Linux runner that has no such hasher at all.
   live_hash="$(hash256 "$JAR")"
-  staged_hash="$(hash256 "$JAR_STAGED")"
+  other_hash="$(hash256 "$other_path")"
   default_result="$(jar_id)"
-  explicit_result="$(jar_id "$JAR_STAGED")"
+  explicit_result="$(jar_id "$other_path")"
-  JAR="$saved_jar"; JAR_STAGED="$saved_staged"
+  JAR="$saved_jar"
-  [ "$live_hash" != "$staged_hash" ] || fail "test fixture error: live and staged jars hashed the same"
+  [ "$live_hash" != "$other_hash" ] || fail "test fixture error: live and other jars hashed the same"
   assert_equals "$live_hash" "$default_result" "jar_id with no arguments must report the hash of \$JAR"
-  assert_equals "$staged_hash" "$explicit_result" "jar_id \"\$JAR_STAGED\" must report the hash of the staged jar, not fall back to \$JAR"
+  assert_equals "$other_hash" "$explicit_result" "jar_id with an explicit path must report the hash of that path, not fall back to \$JAR"
 }
 
 # fleetd #550 — closes a gap the test above leaves open. That test's own reference hash is now ALSO
@@ -651,34 +684,77 @@ test_jar_id_reports_unhashable_when_no_hasher_on_path() {
   assert_equals "unhashable" "$explicit_result" "jar_id (explicit path) with no hasher on PATH must report the same third state"
 }
 
-# fleetd #493 — never build into the path a running process holds. stage_built_jar/swap_staged_jar
+# fleetd #664 — report_jar_state is the --check fix: under the old layout $JAR and the build
-# are exercised directly against real files on disk (not stubs), because the whole point is file
+# output were the same file, so a single "jar on disk" fact covered both. Now they can disagree,
+# and this is the function that is supposed to show that. Agreeing case: two files with IDENTICAL
+# content must print both labels and never warn.
+test_report_jar_state_agrees_when_hashes_match() {
+  local dir saved_build="$BUILD_JAR" saved_jar="$JAR" output
+  dir="$TMP/report-jar-agree"; mkdir -p "$dir"
+  BUILD_JAR="$dir/target-fleetd.jar"; JAR="$dir/run-fleetd.jar"
+  printf 'identical jar bytes' > "$BUILD_JAR"
+  printf 'identical jar bytes' > "$JAR"
+  output="$(report_jar_state)"
+  BUILD_JAR="$saved_build"; JAR="$saved_jar"
+  printf '%s' "$output" | grep -qF 'built jar' \
+    || fail "report_jar_state did not label the built jar"
+  printf '%s' "$output" | grep -qF 'running jar' \
+    || fail "report_jar_state did not label the running jar"
+  if printf '%s' "$output" | grep -qF 'differ'; then
+    fail "report_jar_state warned about a mismatch when both jars have identical content"
+  fi
+}
+
+# The disagreeing case: this is the whole point of the ticket — a built jar that is NOT the
+# running jar must be visibly flagged, not silently printed as two unremarkable facts.
+test_report_jar_state_warns_when_hashes_differ() {
+  local dir saved_build="$BUILD_JAR" saved_jar="$JAR" output
+  dir="$TMP/report-jar-differ"; mkdir -p "$dir"
+  BUILD_JAR="$dir/target-fleetd.jar"; JAR="$dir/run-fleetd.jar"
+  printf 'freshly built jar bytes' > "$BUILD_JAR"
+  printf 'older running jar bytes' > "$JAR"
+  output="$(report_jar_state)"
+  BUILD_JAR="$saved_build"; JAR="$saved_jar"
+  printf '%s' "$output" | grep -qF 'differ' \
+    || fail "report_jar_state did not warn when the built jar and running jar disagree"
+}
+
+# Neither file existing (a fresh checkout, never built or deployed) must report two "absent"
+# facts and never a false mismatch warning — "absent" vs "absent" is agreement, not a diff.
+test_report_jar_state_both_absent_is_not_a_mismatch() {
+  local dir saved_build="$BUILD_JAR" saved_jar="$JAR" output
+  dir="$TMP/report-jar-absent"; mkdir -p "$dir"
+  BUILD_JAR="$dir/no-such-target.jar"; JAR="$dir/no-such-run.jar"
+  output="$(report_jar_state)"
+  BUILD_JAR="$saved_build"; JAR="$saved_jar"
+  printf '%s' "$output" | grep -qF 'absent' \
+    || fail "report_jar_state did not report absent for a missing built/running jar"
+  if printf '%s' "$output" | grep -qF 'differ'; then
+    fail "report_jar_state warned about a mismatch when both jars are simply absent"
+  fi
+}
+
+# Reads JAR and BUILD_JAR exactly as the script sources them, with nothing here assigning
+# either first. JAR must resolve outside $MODULE/target/, and JAR must differ from BUILD_JAR:
+# the daemon's live path and Maven's own build output are never the same file.
+test_jar_and_build_jar_are_sourced_outside_target_and_differ() {
+  source "$ROOT/scripts/redeploy-fleetd.sh"
+  case "$JAR" in
+    "$MODULE"/target/*)
+      fail "\$JAR must not live under \$MODULE/target/ — got $JAR" ;;
+  esac
+  [ "$JAR" != "$BUILD_JAR" ] \
+    || fail "\$JAR and \$BUILD_JAR must not be the same path — got $JAR"
+}
+
+# fleetd #493/#664 — never build into the path a running process holds. swap_staged_jar is
+# exercised directly against real files on disk (not stubs), because the whole point is file
 # behavior (does the content move, does the source disappear, does a failure leave both sides
-# intact) that a stubbed function cannot prove.
+# intact) that a stubbed function cannot prove. stage_built_jar no longer exists: under the #664
-test_stage_built_jar_moves_off_live_path() {
+# layout $BUILD_JAR (target/fleetd.jar) was never the live path, so a build has nothing to be
-  local dir jar staged saved_jar="$JAR" saved_staged="$JAR_STAGED"
+# staged OUT of — swap_staged_jar is called directly against $BUILD_JAR/$JAR (see
-  dir="$TMP/stage-ok"; mkdir -p "$dir"
+# test_swap_ordered_after_wait_and_before_start and the report_jar_state tests below for the rest
-  jar="$dir/fleetd.jar"; staged="$dir/fleetd-new.jar"
+# of that seam).
-  printf 'built jar bytes' > "$jar"
-  JAR="$jar"; JAR_STAGED="$staged"
-  stage_built_jar || fail "stage_built_jar rejected a real build output"
-  JAR="$saved_jar"; JAR_STAGED="$saved_staged"
-  [ ! -f "$jar" ] || fail "stage_built_jar left the jar behind at the live path $jar"
-  [ -f "$staged" ] || fail "stage_built_jar did not create the staged jar at $staged"
-  grep -qF 'built jar bytes' "$staged" || fail "staged jar does not carry the built content"
-}
-
-test_stage_built_jar_dies_when_build_produced_nothing() {
-  local dir output rc=0 saved_jar="$JAR" saved_staged="$JAR_STAGED"
-  dir="$TMP/stage-missing"; mkdir -p "$dir"
-  JAR="$dir/fleetd.jar"; JAR_STAGED="$dir/fleetd-new.jar"
-  output="$(stage_built_jar 2>&1)" || rc=$?
-  JAR="$saved_jar"; JAR_STAGED="$saved_staged"
-  [ "$rc" -ne 0 ] || fail "stage_built_jar accepted a missing build output"
-  printf '%s' "$output" | grep -qF "$dir/fleetd.jar" \
-    || fail "refusal message does not name the missing jar path"
-}
-
 test_swap_staged_jar_moves_staged_onto_live() {
   local dir staged live
   dir="$TMP/swap-ok"; mkdir -p "$dir"
@@ -989,24 +1065,26 @@ test_swap_ordered_after_wait_and_before_start() {
     || fail "swap_if_built (line $swap_line) is not before the start section (line $start_line)"
 }
 
-# fleetd #511: the drain-gate abort message (fired when a build has staged a jar but the operator
+# fleetd #511: the drain-gate abort message (fired when a build has produced a jar but the
-# declines the drain confirmation) used to tell the operator to "Rerun (with or without --no-build)"
+# operator declines the drain confirmation) used to tell the operator to "Rerun (with or without
-# to finish the restart. That is wrong — by the time this message can fire, stage_built_jar has
+# --no-build)" to finish the restart. That is wrong. fleetd #664 changed WHY it is wrong: under
-# already moved the jar off $JAR, so a rerun WITH --no-build hits require_no_build_jar's own refusal
+# the old layout a build emptied the live path immediately, so --no-build's own check refused a
-# ("no jar at $JAR — run without --no-build"). Like test_swap_ordered_after_wait_and_before_start
+# rerun for you; now a build never touches the live path at all, so --no-build would NOT refuse —
-# above, this code path is never reached by sourcing (the SOURCED guard stops before the main flow),
+# it would quietly restart the daemon on the OLD jar and throw away the one just built. Like
-# so the only way to pin its exact wording is to read the source.
+# test_swap_ordered_after_wait_and_before_start above, this code path is never reached by sourcing
+# (the SOURCED guard stops before the main flow), so the only way to pin its exact wording is to
+# read the source.
 test_drain_gate_abort_message_says_no_no_build() {
   local src="$ROOT/scripts/redeploy-fleetd.sh" msg
-  msg="$(grep -A3 -F 'aborted — the running daemon was NOT touched, but the freshly built jar is sitting at' "$src")"
+  msg="$(grep -A6 -F 'aborted — the running daemon was NOT touched, but the freshly built jar is sitting at' "$src")"
   [ -n "$msg" ] || fail "could not find the drain-gate staged-jar abort message in redeploy-fleetd.sh"
   if printf '%s' "$msg" | grep -qF 'with or without --no-build'; then
     fail "abort message still claims a rerun WITH --no-build can finish the restart"
   fi
   printf '%s' "$msg" | grep -qF 'WITHOUT --no-build' \
     || fail "abort message does not tell the operator to rerun without --no-build"
-  printf '%s' "$msg" | grep -qF 'no longer at the live path' \
+  printf '%s' "$msg" | grep -qF 'silently discard' \
-    || fail "abort message does not say why --no-build cannot finish the restart"
+    || fail "abort message does not say that --no-build would silently discard the build just made"
 }
 
 # fleetd #517 — the drain-gate abort branch itself. Before this, the only test of this message was
@@ -1149,19 +1227,22 @@ test_refuse_drain_gate_no_build_staged_absent() {
 #
 # fleetd #555 — the main flow's own call site moved: it used to read
 # `refuse_drain_gate "$DO_BUILD" "$JAR_STAGED"` directly; it now reads
-# `run_drain_gate "$OLD_PID" "$ASSUME_YES" "$DO_BUILD" "$JAR_STAGED"`, and run_drain_gate (tested
+# `run_drain_gate "$OLD_PID" "$ASSUME_YES" "$DO_BUILD" "$BUILD_JAR"` (fleetd #664 renamed the
-# directly below by test_run_drain_gate_*) is what calls refuse_drain_gate with its own local names.
+# fourth argument from $JAR_STAGED to $BUILD_JAR — same role, the not-yet-swapped-in jar — when
-# This grep now pins THAT call site — the thing that would go missing if a future edit deleted the
+# that path stopped being a separate staging file and became target/fleetd.jar itself), and
-# main flow's call to run_drain_gate altogether, the same residual gap #521/#528 already accepted for
+# run_drain_gate (tested directly below by test_run_drain_gate_*) is what calls refuse_drain_gate
-# swap_if_built/refuse_drain_gate (sourcing stops before the main flow runs, so no test in this file
+# with its own local names. This grep now pins THAT call site — the thing that would go missing if
-# can do better than reading the source for this one specific gap).
+# a future edit deleted the main flow's call to run_drain_gate altogether, the same residual gap
+# #521/#528 already accepted for swap_if_built/refuse_drain_gate (sourcing stops before the main
+# flow runs, so no test in this file can do better than reading the source for this one specific
+# gap).
 #
 # The grep ends `|| true`: this file runs under `set -euo pipefail`, so an ABSENT needle would fail
 # the assignment and `set -e` would kill the whole suite before the `[ -n ... ] || fail` guard below
 # ever ran — the exact dead-check shape fleetd #528 also flags as a sweep finding (see the PR body).
 test_run_drain_gate_call_site_present() {
   local src="$ROOT/scripts/redeploy-fleetd.sh" call_line
-  call_line="$(grep -Fn 'run_drain_gate "$OLD_PID" "$ASSUME_YES" "$DO_BUILD" "$JAR_STAGED"' "$src" | head -1 | cut -d: -f1 || true)"
+  call_line="$(grep -Fn 'run_drain_gate "$OLD_PID" "$ASSUME_YES" "$DO_BUILD" "$BUILD_JAR"' "$src" | head -1 | cut -d: -f1 || true)"
   [ -n "$call_line" ] \
     || fail "could not find the main flow's run_drain_gate call site in redeploy-fleetd.sh"
 }
@@ -1269,12 +1350,71 @@ test_run_drain_gate_declined_reply_refuses() {
   source "$ROOT/scripts/redeploy-fleetd.sh"
 }
 
+# Writes a launchd-plist fixture naming jar_path as the ProgramArguments entry after "-jar", so
+# check_jar_path_matches_plist has something real to read back.
+write_launchd_plist_fixture() {
+  local path="$1" jar_path="$2"
+  cat > "$path" <<PLIST
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+  <key>Label</key>
+  <string>test.fixture</string>
+  <key>ProgramArguments</key>
+  <array>
+    <string>/usr/bin/java</string>
+    <string>-jar</string>
+    <string>$jar_path</string>
+    <string>fleetd.yaml</string>
+  </array>
+</dict>
+</plist>
+PLIST
+}
+
+# Agreeing case: a plist whose ProgramArguments names the same jar, resolved, must proceed and
+# say so, never die.
+test_check_jar_path_matches_plist_agrees_ok() {
+  local dir jar plist output rc=0
+  dir="$TMP/jar-path-agree"; mkdir -p "$dir/run"
+  jar="$dir/run/fleetd.jar"
+  plist="$dir/agree.plist"
+  write_launchd_plist_fixture "$plist" "$jar"
+  output="$(check_jar_path_matches_plist "$jar" "$plist" 2>&1)" || rc=$?
+  [ "$rc" -eq 0 ] \
+    || fail "check_jar_path_matches_plist must succeed when the plist names the same jar: $output"
+  printf '%s' "$output" | grep -qF 'jar path check' \
+    || fail "check_jar_path_matches_plist did not print the agreement line: $output"
+}
+
+# The disagreeing case, and the positive control this check exists for: a plist naming a
+# different jar must die, naming both paths.
+test_check_jar_path_matches_plist_mismatch_dies() {
+  local dir jar other_jar plist output rc=0
+  dir="$TMP/jar-path-mismatch"; mkdir -p "$dir/run" "$dir/target"
+  jar="$dir/run/fleetd.jar"
+  other_jar="$dir/target/fleetd.jar"
+  plist="$dir/mismatch.plist"
+  write_launchd_plist_fixture "$plist" "$other_jar"
+  output="$(check_jar_path_matches_plist "$jar" "$plist" 2>&1)" || rc=$?
+  [ "$rc" -ne 0 ] \
+    || fail "check_jar_path_matches_plist must die when the plist names a different jar"
+  printf '%s' "$output" | grep -qF "$jar" \
+    || fail "die message does not name this script's jar path: $output"
+  printf '%s' "$output" | grep -qF "$other_jar" \
+    || fail "die message does not name the plist's jar path: $output"
+}
+
 # fleetd #555 item 4 — the report-state dispatch on $SUPERVISOR_KIND. Inverting this used to report
 # the wrong supervisor and, for the launchd arm specifically, skip check_log_path_matches_plist.
 CHECK_LOG_PATH_CALLED=0
+CHECK_JAR_PATH_CALLED=0
 stub_check_log_path_recorder() {
   CHECK_LOG_PATH_CALLED=0
+  CHECK_JAR_PATH_CALLED=0
   check_log_path_matches_plist() { CHECK_LOG_PATH_CALLED=1; }
+  check_jar_path_matches_plist() { CHECK_JAR_PATH_CALLED=1; }
 }
 
 test_report_supervisor_state_launchd_sets_supervised_and_checks_log_path() {
@@ -1285,6 +1425,8 @@ test_report_supervisor_state_launchd_sets_supervised_and_checks_log_path() {
   [ "$SUPERVISED" = 1 ] || fail "report_supervisor_state launchd must set SUPERVISED=1"
   [ "$CHECK_LOG_PATH_CALLED" = 1 ] \
     || fail "report_supervisor_state launchd must call check_log_path_matches_plist"
+  [ "$CHECK_JAR_PATH_CALLED" = 1 ] \
+    || fail "report_supervisor_state launchd must call check_jar_path_matches_plist"
   source "$ROOT/scripts/redeploy-fleetd.sh"
 }
 
@@ -1296,6 +1438,8 @@ test_report_supervisor_state_systemd_sets_supervised_without_log_path_check() {
   [ "$SUPERVISED" = 1 ] || fail "report_supervisor_state systemd must set SUPERVISED=1"
   [ "$CHECK_LOG_PATH_CALLED" = 0 ] \
     || fail "report_supervisor_state systemd must NOT call check_log_path_matches_plist"
+  [ "$CHECK_JAR_PATH_CALLED" = 0 ] \
+    || fail "report_supervisor_state systemd must NOT call check_jar_path_matches_plist"
   source "$ROOT/scripts/redeploy-fleetd.sh"
 }
 
@@ -2153,6 +2297,8 @@ test_assert_single_daemon_accepts_one_pid
 test_assert_single_daemon_rejects_two_pids
 test_running_pid_excludes_self_matching_wrapper_shell
 test_running_pid_finds_a_real_java_named_second_process
+test_running_pid_finds_a_real_java_named_process_from_target_dir
+test_running_pid_excludes_self_matching_wrapper_shell_naming_target_dir
 test_running_pid_drops_a_pid_whose_comm_is_not_java
 test_running_pid_drops_a_pid_that_exited_before_the_comm_lookup
 test_running_pid_counts_a_pid_whose_comm_is_java
@@ -2161,8 +2307,10 @@ test_jar_id_defaults_to_live_and_reports_explicit_path
 test_hash256_computes_a_real_sha256
 test_jar_id_reports_absent_for_missing_file
 test_jar_id_reports_unhashable_when_no_hasher_on_path
-test_stage_built_jar_moves_off_live_path
+test_report_jar_state_agrees_when_hashes_match
-test_stage_built_jar_dies_when_build_produced_nothing
+test_report_jar_state_warns_when_hashes_differ
+test_report_jar_state_both_absent_is_not_a_mismatch
+test_jar_and_build_jar_are_sourced_outside_target_and_differ
 test_swap_staged_jar_moves_staged_onto_live
 test_swap_staged_jar_dies_without_staged_file
 test_swap_staged_jar_dies_when_mv_fails
@@ -2204,6 +2352,8 @@ test_drain_confirmed_false_on_anything_else
 test_run_drain_gate_skips_prompt_when_not_required
 test_run_drain_gate_confirmed_reply_does_not_refuse
 test_run_drain_gate_declined_reply_refuses
+test_check_jar_path_matches_plist_agrees_ok
+test_check_jar_path_matches_plist_mismatch_dies
 test_report_supervisor_state_launchd_sets_supervised_and_checks_log_path
 test_report_supervisor_state_systemd_sets_supervised_without_log_path_check
 test_report_supervisor_state_none_leaves_supervised_zero