fleetd #234 round 4: invert ExhaustionSink's abstract method so the bug class is unrepresentable

Round 3's factory fixed the two known call sites but the underlying shape
was still there: a lambda written against ExhaustionSink binds to whichever
overload is abstract, and the 2-arg form held that position, so ANY lambda
-- a call-site forwarder, a hand-built test double, a future caller who has
never heard of fleetd #234 -- could still silently take the hint-dropping
default. Two rounds shipped exactly that mistake in two different places.

Fix: made the 3-arg onExhausted(target, reason, profile) the interface's
single abstract method; the 2-arg form is now a default that delegates with
a null profile. A lambda declared against ExhaustionSink today is forced by
the compiler to take three parameters -- there is no overload left for it to
bind to that can drop the hint. This is enforced by the type system, not by
a test that has to remember to check for it.

Knock-on changes:
- ExhaustionSink.none() -- a 3-arg lambda, still a genuine no-op, now safe
  by construction rather than by care.
- ExhaustionSink.forwardingTo(...) -- collapses to a one-line 3-arg lambda;
  kept as a named factory (round 3's lesson: a test must call the real
  object, not rebuild its shape).
- Fleetd.java's real sink and the two OpenCodeLauncherTest sinks that used
  to be anonymous classes overriding both overloads are now plain lambdas
  too -- the 2-arg override each carried was pure boilerplate once the
  interface provides it as a default.
- CompletionResolver.java itself: UNCHANGED, zero diff (confirmed via
  `git diff --stat` before staging) -- its two call sites still call the
  2-arg onExhausted(target, reason), which is now the default and behaves
  identically. CompletionResolverTest (41 tests, 0 failures) proves this;
  its five ExhaustionSink lambdas needed a mechanical third parameter added
  to keep compiling against the new abstract method, no assertion changed.

Mutation proof, re-run against the new shape: forwardingTo's body edited to
call the 2-arg default instead of passing the hint through (the equivalent
of round 3's "delete the 3-arg override" now that there is only one method
to break) -- both new tests go red with the same assertions as round 3:

  ExhaustionSinkForwardingHazardTest...: expected: <gx> but was: <null>
  OpenCodeLauncherTest...ForwardingHop:  expected: <true> but was: <false>
  Tests run: 68, Failures: 2

Restored, re-ran: green (Tests run: 109, Failures: 0, including
CompletionResolverTest).

Compiler proof (not committed -- a scratch file outside the worktree,
compiled with the real ExhaustionSink.java on the classpath, then deleted):

  ExhaustionSink forwarder = (target, reason) -> System.out.println(target + reason);

  error: incompatible types: incompatible parameter types in lambda expression

A 2-arg lambda against this interface no longer compiles at all.

mvn clean install: Tests run: 1129, Failures: 0, Errors: 0, Skipped: 0, BUILD SUCCESS.

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

@@ -169,6 +169,18 @@ public final class Fleetd {
             }
         });
         List<HerdrPeerLauncher> adapters = new ArrayList<>();
+        // fleetd #175: the daemon's real ExhaustionSink can only be built once `sessions` exists
+        // (below), but `sessions` needs `workers`, which needs the adapters built right here — a
+        // genuine cycle. Break it exactly like liveCountRef below: a forwarding sink built now,
+        // pointed at the real one once it exists.
+        AtomicReference<ExhaustionSink> exhaustionSinkRef = new AtomicReference<>(ExhaustionSink.none());
+        // fleetd #234, round 4: routed through the shared ExhaustionSink.forwardingTo factory
+        // rather than written inline here — not because a lambda at this call site is unsafe
+        // anymore (it is not: the 3-arg overload is now the interface's single abstract method, so
+        // there is no 2-arg overload left for any lambda to silently bind to instead), but so a
+        // test can call the exact same object this line builds, instead of asserting a copy of its
+        // shape (round 3's lesson).
+        ExhaustionSink forwardingExhaustionSink = ExhaustionSink.forwardingTo(exhaustionSinkRef::get);
         // The claude-code adapter is the always-present default; keep it even with no profiles (so a
         // bridge configured with no workers, or opencode-only, still has a well-defined base adapter)
         // unless opencode is the only kind configured.
@@ -184,7 +196,7 @@ public final class Fleetd {
                     opencodeProfiles, cfg.effectiveDefaultProfile(), System::getenv,
                     cfg.spawnReadyTimeoutMs(), cfg.spawnReadyPollMs(),
                     () -> config.get().fleet(),
-                    () -> config.get().memberCredentials(), config::get));
+                    () -> config.get().memberCredentials(), config::get, forwardingExhaustionSink));
         }
         AtomicReference<Function<String, Integer>> liveCountRef = new AtomicReference<>(_ -> 0);
         // CB-578 stage B: one quarantine tracker for the whole daemon, shared between the launcher
@@ -334,18 +346,49 @@ public final class Fleetd {
         // CREDENTIAL — not the profile name — so a profile sharing that credential (e.g. two models
         // on one OpenAI account) is refused too, not just the one that happened to report it. Reads
         // the profile config live off `config`, so a credentialId edit is hot: no restart needed.
-        ExhaustionSink exhaustionSink = (target, reason) -> sessions.roster().stream()
-                .filter(session -> target.equals(session.terminalId()))
-                .findFirst()
-                .map(MemberSession::profile)
-                .map(profileName -> config.get().profiles().get(profileName))
-                .ifPresent(profile -> {
-                    String credentialId = profile.effectiveCredentialId();
-                    quarantine.quarantine(credentialId);
-                    log.warn("credential '{}' quarantined for {}s (profile '{}' classified "
-                                    + "BACKEND_EXHAUSTED): {}", credentialId,
-                            cfg.quarantineCooldownSeconds(), profile.profile(), reason);
-                });
+        //
+        // fleetd #175: this sink is now also the quarantine target for OpenCodeLauncher's
+        // model-mismatch check — it needs nothing profile-specific from the caller beyond `target`
+        // (a herdr terminal id) and `reason`, so reusing it here is exactly "the existing
+        // ExhaustionSink path", not a new mechanism.
+        //
+        // fleetd #234: that check fires from SessionAwareHandle.agentSessionId(), which runs during
+        // SessionManager.acquire() BEFORE this session is registered in sessions.roster() — so the
+        // roster-only lookup below used to find nothing, .ifPresent silently no-op'd, and the
+        // ERROR the check had just logged ("quarantining this profile's credential") was a lie:
+        // nothing was quarantined, and nothing said so. Two changes: (1) OpenCodeLauncher now
+        // passes its OWN profile name via ExhaustionSink's 3-arg overload — it already has the
+        // FleetConfig.Profile in hand and does not need the roster at all — used here as a
+        // fallback whenever the roster lookup misses; (2) if a profile still cannot be resolved
+        // (neither the roster nor the hint names a configured one), this logs loudly at ERROR
+        // instead of silently doing nothing — a control that cannot act must say so.
+        // fleetd #234, round 4: the 3-arg overload is now ExhaustionSink's single abstract method,
+        // so this is safely a lambda — there is no separate 2-arg overload left for it to bind to
+        // instead and silently drop profileHint (that was rounds 1-3's whole hazard).
+        ExhaustionSink exhaustionSink = (target, reason, profileHint) -> {
+            String profileName = sessions.roster().stream()
+                    .filter(session -> target.equals(session.terminalId()))
+                    .findFirst()
+                    .map(MemberSession::profile)
+                    .orElse(profileHint);
+            FleetConfig.Profile profile = profileName == null ? null : config.get().profiles().get(profileName);
+            if (profile == null) {
+                log.error("quarantine requested for target '{}' ({}) but no profile could be "
+                                + "resolved — the target is not (yet) in the roster, and {} — "
+                                + "credential NOT quarantined (fleetd #234)",
+                        target, reason,
+                        profileHint == null ? "no profile hint was given"
+                                : "the hinted profile '" + profileHint + "' is not configured");
+                return;
+            }
+            String credentialId = profile.effectiveCredentialId();
+            quarantine.quarantine(credentialId);
+            log.warn("credential '{}' quarantined for {}s (profile '{}'): {}", credentialId,
+                    cfg.quarantineCooldownSeconds(), profile.profile(), reason);
+        };
+        // fleetd #175: point the forwarding sink handed to OpenCodeLauncher above at the real one,
+        // now that `sessions` exists to resolve target -> session -> profile.
+        exhaustionSinkRef.set(exhaustionSink);
         AgentControl agents = router.memberAgents();
         CompletionResolver completion = new CompletionResolver(agents, rendezvous, exhaustedPatterns, exhaustionSink);
         // CB-113: deliver only to an available worker (its MCP is connected), never its boot window.

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

@@ -5,17 +5,70 @@ import dev.ltms.fleet.peer.MemberRole;
 /** Optional session lifecycle hook for live member-slot bindings. */
 public interface MemberLifecycle {
 
+    /** A slot held before a member process starts. */
+    record SlotReservation(String slot, String profile) { }
+
     MemberLifecycle NONE = new MemberLifecycle() {
         @Override
-        public void acquired(MemberRole role, String profile, String terminal) {
+        public MemberRole acquired(MemberRole role, String profile, String terminal) {
+            return role; // no registry configured — nothing to bind against, so the request stands
         }
 
         @Override
         public void released(String terminal) {
         }
+
+        @Override
+        public void requireSlotFor(MemberRole role, String profile) {
+            // no registry configured — nothing to validate against, so nothing is refused
+        }
+
+        @Override
+        public SlotReservation reserve(MemberRole role, String profile) {
+            return null;
+        }
+
+        @Override
+        public boolean bind(SlotReservation reservation, String terminal) {
+            return false;
+        }
+
+        @Override
+        public void release(SlotReservation reservation) {
+        }
     };
 
-    void acquired(MemberRole role, String profile, String terminal);
+    /**
+     * Try to bind a newly spawned {@code terminal} into the role it was granted.
+     *
+     * @return the role this session actually holds: {@code role} unchanged for a role with no
+     *         live slot-binding semantics (dev, reviewer), or when the bind succeeded; a fallback
+     *         role — never {@code role} — when a slot-bound role (architect) could not be bound.
+     *         Callers must record THIS value on the session, never the requested {@code role}, so
+     *         a later roster read never reports a role the session does not hold (CB-619). In
+     *         normal operation this fallback should not happen once a reservation has been bound.
+     *         It remains the honest answer if a caller has no reservation, or if binding a
+     *         reservation unexpectedly fails.
+     */
+    MemberRole acquired(MemberRole role, String profile, String terminal);
 
     void released(String terminal);
+
+    /**
+     * Refuse an acquire before anything spawns when {@code role} requires a live slot binding and
+     * no configured slot carries {@code profile} (CB-619 / fleetd #123). A no-op for a role with
+     * no slot-binding semantics.
+     *
+     * @throws IllegalArgumentException naming the role, the profile, and the pools that do carry it
+     */
+    void requireSlotFor(MemberRole role, String profile);
+
+    /** Reserve a matching slot before launch, or refuse before a charter can be delivered. */
+    SlotReservation reserve(MemberRole role, String profile);
+
+    /** Convert a reservation into a live terminal binding. */
+    boolean bind(SlotReservation reservation, String terminal);
+
+    /** Return an unbound reservation after a failed launch. */
+    void release(SlotReservation reservation);
 }

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

@@ -8,6 +8,7 @@ import org.slf4j.LoggerFactory;
 import java.util.Collections;
 import java.util.HashMap;
 import java.util.LinkedHashMap;
+import java.util.List;
 import java.util.Map;
 import java.util.Objects;
 
@@ -55,8 +56,10 @@ public final class MemberRegistry implements MemberLifecycle {
     }
 
     private final Map<String, Entry> slots;
-    /** Live {@code terminal_id → qualified slot key}; guarded by {@code this}. */
+    /** Live {@code terminal_id → qualified slot key}; guarded by {@code terminalToSlot}. */
     private final Map<String, String> terminalToSlot = new HashMap<>();
+    /** Slot keys held between reservation and the terminal binding. Guarded by terminalToSlot. */
+    private final java.util.Set<String> reservedSlots = new java.util.HashSet<>();
 
     /** Flatten every role pool in {@code fleet} into one registry. Leaders are not members. */
     public MemberRegistry(FleetConfig.Fleet fleet) {
@@ -166,7 +169,7 @@ public final class MemberRegistry implements MemberLifecycle {
             if (existingSlot != null) {
                 return slot.equals(existingSlot); // already this slot (idempotent) or a different one
             }
-            if (terminalToSlot.containsValue(slot)) {
+            if (terminalToSlot.containsValue(slot) || reservedSlots.contains(slot)) {
                 return false; // slot already hosts a terminal — no second one
             }
             terminalToSlot.put(terminal, slot);
@@ -206,19 +209,116 @@ public final class MemberRegistry implements MemberLifecycle {
      *
      * <p>The role check is lifecycle policy. {@link CallerResolver} repeats it when resolving a
      * binding, so a later lifecycle regression cannot turn a worker into an architect.
+     *
+     * <p>CB-619 / fleetd #123: the return value is the role this session actually holds, and the
+     * caller is required to record THAT — never the requested {@code role} — on the session. Before
+     * this fix the caller kept the requested role regardless of whether the bind below succeeded, so
+     * a demoted session's {@code GET /members} row still said {@code "architect"} while
+     * {@code fleet_whoami} (which reads the live binding, not the request) correctly said
+     * {@code "worker"} — three sources of truth that disagreed about one live member, silently.
      */
     @Override
-    public void acquired(MemberRole role, String profile, String terminal) {
+    public MemberRole acquired(MemberRole role, String profile, String terminal) {
         if (role != MemberRole.ARCHITECT || terminal == null || terminal.isBlank()) {
-            return;
+            return role;
         }
         // slotsFor preserves definition order, so duplicate-profile slots use the first free one.
         for (Entry entry : slotsFor(MemberRole.ARCHITECT).values()) {
             if (Objects.equals(profile, entry.profile()) && bind(entry.key(), terminal)) {
-                return;
+                return MemberRole.ARCHITECT;
+            }
+        }
+        // fleetd #123: at least WARN — a role downgrade that the roster must now also reflect is
+        // not routine bookkeeping. requireSlotFor already refuses the config-gap case (no slot at
+        // all carries this profile) before a process ever spawns; reaching here means the config DID
+        // carry a matching slot but every one of them was already bound to a different terminal — a
+        // race this pre-spawn check cannot close on its own (see requireSlotFor's javadoc).
+        log.warn("member slot: no free architect slot for profile={} terminal={}; holding the session "
+                + "as {} instead of the architect it asked for — every configured slot for this "
+                + "profile is already bound to a different terminal", profile, terminal,
+                MemberRole.DEV.wireName());
+        return MemberRole.DEV;
+    }
+
+    /**
+     * CB-619 / fleetd #123: refuse an architect acquire before anything spawns when no configured
+     * slot carries {@code profile} — the config-gap case from the original defect report (a spawn
+     * asked for {@code role=architect, profile=sonnet}, and {@code fleet.architects} carried only
+     * {@code opus} and {@code sol}). A dev/reviewer acquire is always a no-op: those pools are
+     * placement candidates only (see {@code CompositePeerLauncher}), never a live identity binding,
+     * so there is nothing here to refuse — an explicit profile outside the pool for those roles is a
+     * documented operator override, not a defect.
+     *
+     * <p>This closes the config-gap case, not the live-capacity case: a profile that DOES carry a
+     * slot can still lose the race to a concurrent spawn between this check and the actual
+     * {@link #bind}, which is why {@link #acquired} must still answer honestly even after this
+     * check has passed.
+     */
+    @Override
+    public void requireSlotFor(MemberRole role, String profile) {
+        if (role != MemberRole.ARCHITECT) {
+            return;
+        }
+        boolean hasSlot = slotsFor(MemberRole.ARCHITECT).values().stream()
+                .anyMatch(e -> Objects.equals(profile, e.profile()));
+        if (hasSlot) {
+            return;
+        }
+        List<String> pools = slotsFor(MemberRole.ARCHITECT).values().stream()
+                .map(Entry::profile)
+                .distinct()
+                .toList();
+        throw new IllegalArgumentException(
+                "no " + role.wireName() + " slot for profile '" + profile + "' — an architect's "
+                        + "identity IS the slot it is bound to, so there is nothing to bind this "
+                        + "session's identity to. fleet." + role.configKey() + " carries profiles: "
+                        + (pools.isEmpty() ? "(none configured)" : String.join(", ", pools))
+                        + "; add profile '" + profile + "' there, or spawn " + role.wireName()
+                        + " on one of those profiles instead");
+    }
+
+    @Override
+    public SlotReservation reserve(MemberRole role, String profile) {
+        if (role != MemberRole.ARCHITECT) {
+            return null;
+        }
+        synchronized (terminalToSlot) {
+            for (Entry entry : slotsFor(MemberRole.ARCHITECT).values()) {
+                if ((profile == null || profile.isBlank() || Objects.equals(profile, entry.profile()))
+                        && !terminalToSlot.containsValue(entry.key()) && reservedSlots.add(entry.key())) {
+                    return new SlotReservation(entry.key(), entry.profile());
+                }
+            }
+        }
+        throw new IllegalArgumentException("no free architect slot for profile '" + profile
+                + "' — every matching slot is already bound or reserved");
+    }
+
+    @Override
+    public boolean bind(SlotReservation reservation, String terminal) {
+        if (reservation == null || terminal == null || terminal.isBlank()) {
+            return false;
+        }
+        synchronized (terminalToSlot) {
+            if (!reservedSlots.remove(reservation.slot())) {
+                return false;
+            }
+            if (!isSlot(reservation.slot()) || terminalToSlot.containsKey(terminal)
+                    || terminalToSlot.containsValue(reservation.slot())) {
+                return false;
+            }
+            terminalToSlot.put(terminal, reservation.slot());
+            return true;
+        }
+    }
+
+    @Override
+    public void release(SlotReservation reservation) {
+        if (reservation != null) {
+            synchronized (terminalToSlot) {
+                reservedSlots.remove(reservation.slot());
             }
         }
-        log.info("member slot: no free architect slot for profile={}; session remains a worker", profile);
     }
 
     /** Unbind a released terminal using the compare-safe registry operation. */

fleetd/src/main/java/dev/ltms/fleet/inject/ExhaustionSink.java

@@ -1,5 +1,7 @@
 package dev.ltms.fleet.inject;
 
+import java.util.function.Supplier;
+
 /**
  * Notified when {@link CompletionResolver} actually delivers a {@code BACKEND_EXHAUSTED}
  * classification to a waiting send (CB-578 stage B) — never on a race that lost (see
@@ -10,22 +12,81 @@ package dev.ltms.fleet.inject;
  * profiles or credentials, so mapping {@code target} to whatever should be quarantined is entirely
  * the sink's job — see {@code Fleetd.main}'s wiring, which resolves target → session → profile →
  * {@code effectiveCredentialId()} and calls {@code BackendQuarantine.quarantine} on it.
+ *
+ * <p><strong>The 3-arg overload is the single abstract method — fleetd #234, round 4.</strong> Two
+ * earlier rounds each shipped a caller that silently dropped the profile hint (see {@link
+ * #onExhausted(String, String, String)}): a lambda written against this interface can only ever
+ * implement whichever overload is abstract, and while the 2-arg form held that position, EVERY
+ * lambda site — a call-site forwarder in {@code Fleetd.java}, a hand-built test double — bound to
+ * it and silently inherited the profile-dropping default, whether or not its author remembered the
+ * hazard. Making the 3-arg form abstract instead removes the shape entirely: a lambda declared
+ * against this interface today is <em>forced</em> by the compiler to take {@code (target, reason,
+ * profile)}, so there is no overload left for it to bind to that can drop the hint. This is a type
+ * change, not a test — it holds even for a caller that has never heard of fleetd #234.
  */
 @FunctionalInterface
 public interface ExhaustionSink {
 
     /**
-     * @param target the herdr terminal id whose turn was classified {@code BACKEND_EXHAUSTED}
-     * @param reason the matched-line reason carried by the classification
+     * @param target  the herdr terminal id whose turn was classified {@code BACKEND_EXHAUSTED}
+     * @param reason  the matched-line reason carried by the classification
+     * @param profile the profile the caller already knows should be quarantined, or {@code null}
+     *                when the caller has no better answer than {@code target} alone (fleetd #234):
+     *                a caller whose {@code target} is not yet resolvable through whatever roster
+     *                the sink's implementation consults — {@link
+     *                dev.ltms.fleet.member.OpenCodeLauncher}'s model-mismatch check fires from
+     *                {@code SessionAwareHandle.agentSessionId()}, which runs during {@code
+     *                SessionManager.acquire()} <em>before</em> that session is registered, so a
+     *                target -> session -> profile lookup finds nothing at that point. That launcher
+     *                already has its own {@code FleetConfig.Profile} in hand and does not need the
+     *                roster to know which profile to quarantine, so it supplies this directly
+     *                instead of leaving the sink to guess.
      */
-    void onExhausted(String target, String reason);
+    void onExhausted(String target, String reason, String profile);
+
+    /**
+     * Convenience for a caller with no profile to offer — every existing call site that predates
+     * the hint (fleetd #234): {@link dev.ltms.fleet.inject.CompletionResolver}'s two call sites
+     * always call with a {@code target} that IS live in the roster at the time of the call, so they
+     * need no hint and keep working exactly as before, unchanged by this default.
+     *
+     * @param target as {@link #onExhausted(String, String, String)}
+     * @param reason as {@link #onExhausted(String, String, String)}
+     */
+    default void onExhausted(String target, String reason) {
+        onExhausted(target, reason, null);
+    }
 
     /**
      * Inert sink — nothing happens on exhaustion. The explicit stand-in a caller (or a test not
      * exercising this feature) passes instead of a defaulting overload, exactly like
-     * {@link ExhaustedPatternLookup#none()}.
+     * {@link ExhaustedPatternLookup#none()}. Safe as a lambda: the 3-arg form is now the interface's
+     * single abstract method, so a lambda here has no other overload to silently bind to instead —
+     * it simply does nothing with all three arguments.
      */
     static ExhaustionSink none() {
-        return (target, reason) -> { };
+        return (target, reason, profile) -> { };
+    }
+
+    /**
+     * A sink that forwards to whatever {@code target} currently supplies (fleetd #234). Exists to
+     * break a genuine construction-order cycle: {@code Fleetd.main} builds its adapters (including
+     * {@link dev.ltms.fleet.member.OpenCodeLauncher}) before {@code sessions} exists, so it cannot
+     * hand them the real sink yet — it hands them a forwarder pointed at an {@code
+     * AtomicReference<ExhaustionSink>} that starts at {@link #none()} and gets {@code .set()} to the
+     * real sink once {@code sessions} is built. {@code target} is evaluated on every call, never
+     * cached, so the forwarder keeps working after the reference is repointed.
+     *
+     * <p>Now safe as a one-line lambda (round 4): forwarding the single 3-arg abstract method
+     * forwards everything a caller can supply — there is no separate 2-arg overload left for a
+     * forwarder to bind to instead and silently lose the hint. Kept as a named factory rather than
+     * written inline at each call site anyway, so a test can call the exact object {@code
+     * Fleetd.java} builds instead of asserting a rebuilt copy of its shape (round 3's lesson).
+     *
+     * @param target supplies the sink to forward to, evaluated fresh on every call
+     * @return a sink whose call delegates to {@code target.get()}
+     */
+    static ExhaustionSink forwardingTo(Supplier<ExhaustionSink> target) {
+        return (t, r, p) -> target.get().onExhausted(t, r, p);
     }
 }

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

@@ -459,12 +459,83 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
      * {@link OpenCodeLauncher#writeConfig} already uses for its charter file, since the process that
      * reads this file (the spawned peer) outlives this JVM call and there is no spawn-scoped teardown
      * hook to delete it synchronously.
+     *
+     * <p><b>fleetd #222.</b> {@code Files.createTempFile(prefix, suffix)} with no directory argument
+     * resolves against {@code java.io.tmpdir} — on macOS the per-user {@code $TMPDIR} under
+     * {@code /var/folders/...}, mode {@code 0700}, both resolved against FLEETD's own OS user. Under
+     * {@code memberHerdrSocket:} the member pane runs as a DIFFERENT OS user, so that user cannot even
+     * traverse the directory, let alone read the file — and since fleetd #220 the charter file is the
+     * ONLY delivery path for {@code --append-system-prompt-file}, always, not merely the fallback it
+     * used to be. A member handed a path it cannot read is not degraded, it is broken: see this
+     * method's refusal branch below.
+     *
+     * <p><b>Measured severity (fleetd #222 real-binary check, claude 2.1.258):</b> an unreadable
+     * {@code --append-system-prompt-file} is the LOUD failure, not the silent one. {@code claude}
+     * checks the file before touching auth or the network — invoked with a bogus API key against a
+     * {@code chmod 000} file, it printed {@code Error reading append system prompt file: EACCES:
+     * permission denied, open '<path>'} and exited 1 immediately (a nonexistent path gets {@code
+     * Error: Append system prompt file not found: <path>}, same exit code). So the pre-fix bug did
+     * NOT leave a charter-less member silently occupying a pane and never calling {@code
+     * fleet_reply} — it made the herdr pane exit immediately, which the CB-306 spawn-readiness gate
+     * (this launcher's {@code spawnReadyTimeoutMs} poll) would have surfaced as "did not reach
+     * injectable state", the same unexplained-timeout shape fleetd #220 already describes. Still a
+     * real defect (every claude-code member under {@code memberHerdrSocket} would have failed to
+     * spawn), but not the worse, undetectable failure mode.
+     *
+     * <ul>
+     *   <li>{@code memberHerdrSocket} ABSENT (today's only live mode): byte-identical to before this
+     *       fix — {@code Files.createTempFile("fleetd-role-charter-", ".md")} with no directory
+     *       argument, i.e. still resolved against {@code java.io.tmpdir}.</li>
+     *   <li>{@code memberHerdrSocket} PRESENT: a fresh per-spawn directory is created under {@code
+     *       worktreeRoot} (never {@code java.io.tmpdir}) holding just the charter file, then shared
+     *       read-only with {@code worktreeGroup} via {@link EnvAllowListScrub#shareWithGroup} — the
+     *       SAME mechanism fleetd #213 built for the ZDOTDIR scrub and fleetd #219 reused for {@link
+     *       OpenCodeLauncher#writeConfig}'s {@code opencode.json} directory, reused here rather than
+     *       duplicated a third time. A per-spawn subdirectory (not {@code worktreeRoot} itself) is the
+     *       unit {@code shareWithGroup} chmods, so this never touches permissions on anything else
+     *       under {@code worktreeRoot}.</li>
+     * </ul>
+     *
+     * <p><b>Follows fleetd #219's REFUSAL decision, not #213's degrade decision.</b> The ZDOTDIR
+     * scrub is a credential CONTROL — a degraded control (the CB-596 sentinel overlay) still has
+     * value, so #213 falls back rather than refusing. A charter is NOT a control, it is the member's
+     * TURN CONTRACT (the rule that ends every turn with {@code fleet_reply}). A member spawned with no
+     * charter is not degraded, it is broken: either claude-code exits on the unreadable
+     * {@code --append-system-prompt-file} path and the spawn dies at the readiness gate (loud), or it
+     * starts anyway with no charter and never calls {@code fleet_reply} — the sender silently gets
+     * nothing (silent). Neither outcome is worth trading for "spawn something." So a missing {@code
+     * worktreeRoot}/{@code worktreeGroup} under {@code memberHerdrSocket} refuses the spawn here,
+     * naming the missing key, exactly like {@link OpenCodeLauncher#configParentDir()}.
+     *
+     * @throws IllegalStateException when {@code memberHerdrSocket} is configured but {@code
+     *                                worktreeRoot} and/or {@code worktreeGroup} is not
      */
-    private static Path writeCharterFile(String charterText) {
+    private Path writeCharterFile(String charterText) {
         try {
-            Path file = Files.createTempFile("fleetd-role-charter-", ".md");
+            if (!memberHerdrSocketConfigured()) {
+                Path file = Files.createTempFile("fleetd-role-charter-", ".md");
+                Files.writeString(file, charterText);
+                file.toFile().deleteOnExit();
+                return file;
+            }
+            Path parentDir = memberScrubParentDir();
+            String group = memberGroup();
+            if (parentDir == null || group == null) {
+                throw new IllegalStateException("memberHerdrSocket is configured, so the role/reply "
+                        + "charter file (mounted via --append-system-prompt-file) must be placed where "
+                        + "the member's OS user can read it — worktreeRoot, shared via worktreeGroup — "
+                        + "but " + (parentDir == null ? "worktreeRoot" : "worktreeGroup") + " is not "
+                        + "configured. Refusing to spawn rather than hand the member a charter path it "
+                        + "cannot read: that member's turn contract (the fleet_reply rule) would never "
+                        + "reach it. Configure both worktreeRoot and worktreeGroup to enable claude-code "
+                        + "member spawns under memberHerdrSocket.");
+            }
+            Path dir = Files.createTempDirectory(parentDir, "fleetd-role-charter-");
+            dir.toFile().deleteOnExit();
+            Path file = dir.resolve("charter.md");
             Files.writeString(file, charterText);
             file.toFile().deleteOnExit();
+            EnvAllowListScrub.shareWithGroup(dir, group);
             return file;
         } catch (IOException e) {
             throw new UncheckedIOException("cannot write role charter temp file", e);

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

@@ -3,11 +3,13 @@ package dev.ltms.fleet.member;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.herdr.Agent;
 import dev.ltms.fleet.herdr.AgentControl;
+import dev.ltms.fleet.herdr.AgentStatus;
 import dev.ltms.fleet.herdr.HerdrClient;
 import dev.ltms.fleet.herdr.HerdrException;
 import dev.ltms.fleet.herdr.Tab;
 import dev.ltms.fleet.herdr.Workspace;
 import dev.ltms.fleet.herdr.WorkspaceControl;
+import dev.ltms.fleet.inject.StatusRefiner;
 import dev.ltms.fleet.peer.Capability;
 import dev.ltms.fleet.peer.CharterReceipt;
 import dev.ltms.fleet.peer.MemberRole;
@@ -151,6 +153,14 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
     private final LongSupplier nowMillis; // monotonic clock (injectable for tests)
     private final Runnable sleeper; // sleep/wait hook (injectable for tests; never real-sleep in unit tests)
 
+    /**
+     * fleetd #176 fix 2: the same UNKNOWN-refinement {@code dev.ltms.fleet.inject.StatusPoller}
+     * uses, reused here for the spawn-readiness gate. Constructed once from {@link #agents} — see
+     * {@link #refinedInjectable(String, Agent)} for the corroboration that keeps it from firing on
+     * a dead pane's bare shell prompt.
+     */
+    private final StatusRefiner statusRefiner;
+
     // Per-process token mixed into each peer name so a fresh process (nameSeq back at 0) cannot
     // collide with same-profile peers that outlived a restart. See startUniquelyNamed.
     private final String nameNonce = String.format("%06x", new SecureRandom().nextInt(1 << 24));
@@ -272,6 +282,7 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         this.memberCredentials = memberCredentials;
         this.hostEnvNames = hostEnvNames != null ? hostEnvNames : () -> System.getenv().keySet();
         this.config = config;
+        this.statusRefiner = new StatusRefiner(agents);
     }
 
     // --- adapter seams -------------------------------------------------------------------------
@@ -909,16 +920,35 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
     // --- spawn-readiness gate (CB-306) ---------------------------------------------------------
 
     /**
-     * Poll {@link AgentControl#status} until the pane reports an injectable state or the configured
+     * Poll {@link AgentControl#get} until the pane reports an injectable state or the configured
      * timeout elapses. On timeout, close the pane (self-reap) and throw.
+     *
+     * <p>fleetd #176 fix 1: {@code agents.get} was previously unguarded here, so a herdr
+     * {@code *_not_found} answer — which is what happens when the backend process EXITED rather
+     * than being slow — propagated as a raw {@link HerdrException} instead of the
+     * {@link PeerUnreachableException} every other failure path of this gate produces, and skipped
+     * teardown ({@link #stop}) entirely, leaking the pane/tab. {@link #failFastOnGoneBackend} closes
+     * that gap: it stops waiting immediately (never burns the rest of the timeout), runs the same
+     * teardown the timeout path below runs, and throws with a message that says the backend exited
+     * rather than that the pane was slow. Any other {@link HerdrException} still propagates
+     * unchanged — this gate does not know how to recover from it.
      */
     private void waitUntilInjectableOrThrow(String paneId) {
-        long deadline = nowMillis.getAsLong() + spawnReadyTimeoutMs;
-        Object lastStatus = null;
+        long start = nowMillis.getAsLong();
+        long deadline = start + spawnReadyTimeoutMs;
+        AgentStatus lastStatus = null;
         while (nowMillis.getAsLong() < deadline) {
-            var status = agents.status(paneId);
-            lastStatus = status;
-            if (status.injectable()) {
+            Agent sample;
+            try {
+                sample = agents.get(paneId);
+            } catch (HerdrException e) {
+                if (isAlreadyGone(e)) {
+                    failFastOnGoneBackend(paneId, e, nowMillis.getAsLong() - start);
+                }
+                throw e; // any other herdr failure is not ours to interpret — let it propagate
+            }
+            lastStatus = sample.status();
+            if (lastStatus.injectable() || refinedInjectable(paneId, sample)) {
                 log.debug("peer pane={} reached injectable state", paneId);
                 return;
             }
@@ -937,6 +967,66 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
                         + spawnReadyTimeoutMs + "ms");
     }
 
+    /**
+     * fleetd #176 fix 1: the backend process exited while the gate was still waiting — herdr
+     * answered {@code *_not_found} instead of ever reporting an injectable status. Fails
+     * immediately (never burns the rest of {@link #spawnReadyTimeoutMs}), runs the exact same
+     * teardown {@link #waitUntilInjectableOrThrow}'s timeout path runs, and throws a
+     * {@link PeerUnreachableException} whose message says the process exited rather than that the
+     * pane was slow.
+     *
+     * @throws PeerUnreachableException always — this method never returns normally
+     */
+    private void failFastOnGoneBackend(String paneId, HerdrException cause, long elapsedMs) {
+        String tail = readPaneQuietly(paneId); // read before stop() closes the pane
+        log.warn("peer pane={} backend process exited after {}ms while waiting for injectable "
+                        + "state (herdr: {}) — closing. Pane tail:\n{}",
+                paneId, elapsedMs, cause.getMessage(), tail);
+        stop(paneId);
+        throw new PeerUnreachableException(
+                "worker pane " + paneId + " backend process exited after " + elapsedMs
+                        + "ms while waiting to become injectable (herdr reported: " + cause.getMessage()
+                        + "). Pane tail:\n" + tail);
+    }
+
+    /**
+     * fleetd #176 fix 2: resolve a raw {@link AgentStatus#UNKNOWN} sample into a trustworthy
+     * injectable state via pane content, the same refinement {@code StatusPoller} applies during a
+     * peer's working life — but corroborated, because the trap this gate is exposed to that the
+     * poller is not: a pane whose backend has already exited settles at a plain shell prompt, and
+     * that prompt commonly contains the same {@code ❯} glyph {@link StatusRefiner#classify} treats
+     * as "idle at the Claude Code TUI prompt". Naively wiring the refiner in would turn "the backend
+     * died" into "ready to inject" — strictly worse than today's timeout.
+     *
+     * <p>Two guards, both required:
+     * <ul>
+     *   <li>{@link StatusRefiner#classify} is written for the Claude Code TUI only (its own javadoc
+     *       says so), so refinement only ever runs for the {@code claude} adapter — never for
+     *       {@code opencode} or any future non-Claude backend, whatever its pane looks like.</li>
+     *   <li>The refined status is accepted only when the <em>same</em> {@code agents.get} sample
+     *       ({@code sample}, taken once by the caller) still reports a non-null
+     *       {@link Agent#agentType()} — herdr's own "a supported backend is still detected here"
+     *       signal, read from the very sample the raw status came from so the two can never
+     *       disagree. A bare shell prompt left by an exited backend reports no {@code agentType}.</li>
+     * </ul>
+     *
+     * <p>Called only when {@code sample.status() == UNKNOWN}, so a healthy spawn — which never sees
+     * {@code UNKNOWN} — triggers zero extra herdr calls; only a persistently-{@code unknown} pane
+     * pays the one extra {@code agent.read} {@link StatusRefiner#refine} performs.
+     */
+    private boolean refinedInjectable(String paneId, Agent sample) {
+        if (sample.status() != AgentStatus.UNKNOWN) {
+            return false;
+        }
+        if (!"claude".equals(namePrefix)) {
+            return false; // StatusRefiner.classify reads a Claude Code TUI prompt specifically
+        }
+        if (sample.agentType() == null) {
+            return false; // no corroborating liveness signal — could be a dead pane's bare shell
+        }
+        return statusRefiner.refine(paneId, AgentStatus.UNKNOWN, agents).injectable();
+    }
+
     /**
      * fleetd #220: the pane's recent output, clipped, for the readiness-gate timeout log — or a
      * short note when it cannot be read. Best-effort by construction: this runs on a path that is

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

@@ -6,6 +6,7 @@ import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.herdr.Agent;
 import dev.ltms.fleet.herdr.AgentControl;
 import dev.ltms.fleet.herdr.WorkspaceControl;
+import dev.ltms.fleet.inject.ExhaustionSink;
 import dev.ltms.fleet.peer.Capability;
 import dev.ltms.fleet.peer.CharterReceipt;
 import dev.ltms.fleet.peer.PeerHandle;
@@ -21,6 +22,7 @@ import java.util.List;
 import java.util.Map;
 import java.util.Set;
 import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.BooleanSupplier;
 import java.util.function.Function;
 import java.util.function.LongSupplier;
@@ -73,6 +75,19 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
      */
     private final OpenCodeSessionDiscovery discovery;
 
+    /**
+     * fleetd #175: notified when {@link SessionAwareHandle} detects, on the same late-resolve
+     * read that discovers the session id, that the live opencode session is running a DIFFERENT
+     * model than the profile requested — opencode does not fail on an unknown {@code -m}, it
+     * silently falls back to a default (potentially paid) model. Reused exactly as
+     * {@code CompletionResolver}'s {@code BACKEND_EXHAUSTED} path uses it: this launcher supplies
+     * only the herdr terminal id and a reason string; mapping target → session → profile →
+     * credential stays entirely the sink's job (see {@code Fleetd.main}'s wiring). Defaults to
+     * {@link ExhaustionSink#none()} for a caller (an older constructor, or a test not exercising
+     * this) that opts out.
+     */
+    private final ExhaustionSink exhaustionSink;
+
     /**
      * Production constructor — disables the spawn-ready gate ({@code spawnReadyTimeoutMs == 0}) so it
      * matches the legacy non-blocking spawn semantics. Config dirs are created under the JVM temp dir.
@@ -132,9 +147,25 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                              Supplier<FleetConfig.Fleet> fleet,
                              Supplier<FleetConfig.MemberCredentials> memberCredentials,
                              Supplier<FleetConfig> config) {
+        this(agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs, spawnReadyPollMs,
+                fleet, memberCredentials, config, ExhaustionSink.none());
+    }
+
+    /**
+     * Production constructor, plus the live config for URI environment exclusions and the fleetd
+     * #175 model-mismatch quarantine sink.
+     */
+    public OpenCodeLauncher(AgentControl agents, WorkspaceControl spaces,
+                             Map<String, FleetConfig.Profile> profiles, String defaultProfile,
+                             Function<String, String> env, long spawnReadyTimeoutMs, long spawnReadyPollMs,
+                             Supplier<FleetConfig.Fleet> fleet,
+                             Supplier<FleetConfig.MemberCredentials> memberCredentials,
+                             Supplier<FleetConfig> config,
+                             ExhaustionSink exhaustionSink) {
         this(agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs,
                 System::currentTimeMillis, () -> sleepUninterruptibly(spawnReadyPollMs),
-                defaultConfigRoot(), defaultDiscoveryRoot(), fleet, memberCredentials, config);
+                defaultConfigRoot(), defaultDiscoveryRoot(), fleet, memberCredentials, config,
+                exhaustionSink);
     }
 
     /**
@@ -182,6 +213,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                 spawnReadyTimeoutMs, nowMillis, sleeper, fleet);
         this.configRoot = configRoot;
         this.discovery = new OpenCodeSessionDiscovery(discoveryRoot);
+        this.exhaustionSink = ExhaustionSink.none();
     }
 
     /**
@@ -207,10 +239,29 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                              Supplier<FleetConfig.Fleet> fleet,
                              Supplier<FleetConfig.MemberCredentials> memberCredentials,
                              Supplier<FleetConfig> config) {
+        this(agents, spaces, profiles, defaultProfile, env, spawnReadyTimeoutMs, nowMillis, sleeper,
+                configRoot, discoveryRoot, fleet, memberCredentials, config, ExhaustionSink.none());
+    }
+
+    /**
+     * Full testability constructor, plus the live config for URI environment exclusions and the
+     * fleetd #175 model-mismatch quarantine sink — the constructor a test drives directly to
+     * observe {@link ExhaustionSink#onExhausted} without going through {@code Fleetd.main}'s
+     * wiring.
+     */
+    public OpenCodeLauncher(AgentControl agents, WorkspaceControl spaces,
+                             Map<String, FleetConfig.Profile> profiles, String defaultProfile,
+                             Function<String, String> env, long spawnReadyTimeoutMs,
+                             LongSupplier nowMillis, Runnable sleeper, Path configRoot, Path discoveryRoot,
+                             Supplier<FleetConfig.Fleet> fleet,
+                             Supplier<FleetConfig.MemberCredentials> memberCredentials,
+                             Supplier<FleetConfig> config,
+                             ExhaustionSink exhaustionSink) {
         super(NAME_PREFIX, agents, spaces, profiles, defaultProfile, env,
                 spawnReadyTimeoutMs, nowMillis, sleeper, fleet, memberCredentials, null, config);
         this.configRoot = configRoot;
         this.discovery = new OpenCodeSessionDiscovery(discoveryRoot);
+        this.exhaustionSink = exhaustionSink == null ? ExhaustionSink.none() : exhaustionSink;
     }
 
     private static Path defaultConfigRoot() {
@@ -622,8 +673,13 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
     @Override
     public PeerHandle spawn(SpawnRequest req) {
         PeerHandle inner = super.spawn(req);
-        return new SessionAwareHandle(inner, discovery, effectiveCwd(req),
-                this::memberHerdrSocketConfigured, discoveryUnavailableWarned);
+        // fleetd #175: the same profile config buildLaunch resolved for this spawn (requireProfile
+        // is deterministic on req.profileName(), so re-resolving here costs a map lookup, not a
+        // second decision) — SessionAwareHandle needs cfg.model() to know what THIS session should
+        // be running.
+        FleetConfig.Profile cfg = requireProfile(req.profileName());
+        return new SessionAwareHandle(inner, discovery, effectiveCwd(req), cfg,
+                this::memberHerdrSocketConfigured, discoveryUnavailableWarned, exhaustionSink);
     }
 
     /**
@@ -633,22 +689,48 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
      * are untouched — only the opencode-specific identity answer is added. {@code sessionName()}
      * stays null: opencode has no display-name seam, so the logical name lives only in the bridge's
      * roster (see the SESSION_NAME capability).
+     *
+     * <p>fleetd #175: {@link #agentSessionId()} is also where the model-mismatch check lives (see
+     * {@link #checkModelMatch()}) — it is the one method the real late-resolve path
+     * ({@code SessionManager.resolveAgentSessionId}, via {@code get}/{@code rosterResolved}/
+     * {@code release}) actually calls, and only while the session id is still unknown. Putting the
+     * check anywhere else risks repeating PR #203's mistake: a check that runs before opencode has
+     * written the row it needs, and so never fires.
      */
     private static final class SessionAwareHandle implements PeerHandle {
         private final PeerHandle delegate;
         private final OpenCodeSessionDiscovery discovery;
         private final String cwd;
+        private final FleetConfig.Profile cfg;
         private final BooleanSupplier discoveryUnavailable;
         private final AtomicBoolean discoveryUnavailableWarned;
+        private final ExhaustionSink exhaustionSink;
+        /** CAS'd true the first (and only) time a model mismatch is reported for this handle. */
+        private final AtomicBoolean modelMismatchReported = new AtomicBoolean();
+        /**
+         * The session id, once {@link OpenCodeSessionDiscovery#sessionIdForDirectory} first
+         * resolves a non-null answer for this handle (fleetd #234). Sticky on purpose: {@code
+         * directory} is a shared-cwd heuristic (see {@link OpenCodeSessionDiscovery}'s class
+         * javadoc) that can start returning a DIFFERENT row once another session shares the same
+         * directory and writes a newer one — re-deriving it on every call would let this handle's
+         * identity silently drift to a sibling's session. Once resolved, this IS the answer, and
+         * {@link #checkModelMatch} reads only the row this id names, never "whatever is newest in
+         * the directory right now."
+         */
+        private final AtomicReference<String> resolvedSessionId = new AtomicReference<>();
 
         SessionAwareHandle(PeerHandle delegate, OpenCodeSessionDiscovery discovery, String cwd,
+                            FleetConfig.Profile cfg,
                             BooleanSupplier discoveryUnavailable,
-                            AtomicBoolean discoveryUnavailableWarned) {
+                            AtomicBoolean discoveryUnavailableWarned,
+                            ExhaustionSink exhaustionSink) {
             this.delegate = delegate;
             this.discovery = discovery;
             this.cwd = cwd;
+            this.cfg = cfg;
             this.discoveryUnavailable = discoveryUnavailable;
             this.discoveryUnavailableWarned = discoveryUnavailableWarned;
+            this.exhaustionSink = exhaustionSink;
         }
 
         @Override
@@ -689,11 +771,93 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                 }
                 return null;
             }
+            // fleetd #234: once resolved, stay resolved. Re-deriving from `directory` on every call
+            // would let this handle's identity drift to a sibling session that later shares the
+            // same cwd and writes a newer row — see resolvedSessionId's javadoc.
+            String cached = resolvedSessionId.get();
+            if (cached != null) {
+                return cached;
+            }
             // Lazy + retried, never a spawn-time blocker: opencode writes the session record only
             // when the session is first persisted, so null here is the correct interim answer and
             // the caller re-calls later (each call re-scans, picking up a record that has since
             // appeared).
-            return discovery.sessionIdForDirectory(cwd);
+            String id = discovery.sessionIdForDirectory(cwd);
+            if (id != null) {
+                resolvedSessionId.compareAndSet(null, id);
+            }
+            // fleetd #175: check on the SAME tick — while the caller (SessionManager's late-resolve
+            // step) is still re-polling because the id is unknown, the row this id came from (once
+            // it exists) is exactly the row that also carries the actual model. Once id resolves,
+            // the caller stops calling agentSessionId() for this session, so this is naturally a
+            // once-only check that happens right when the row first appears.
+            checkModelMatch(id);
+            return id;
+        }
+
+        /**
+         * Verify the live opencode session is running the model {@link #cfg} requested (fleetd
+         * #175) and, on a real mismatch, log an ERROR and quarantine through {@link
+         * #exhaustionSink}. A no-op when there is nothing to compare against — no model configured,
+         * already reported once for this handle, {@code sessionId} itself is not resolved yet
+         * (fleetd #234: absent evidence, not a mismatch), or the actual model is still UNKNOWN (no
+         * row yet, unreadable database, or unparseable evidence). UNKNOWN must never be treated as
+         * a mismatch: that is the single most important safety rule here — a false positive would
+         * quarantine a perfectly working profile's credential.
+         *
+         * @param sessionId the id {@link #agentSessionId()} just resolved (or had cached) for THIS
+         *                  handle — the model is read back for this exact session (fleetd #234's
+         *                  {@link OpenCodeSessionDiscovery#actualModelForSessionId}), never
+         *                  re-derived from {@code directory}
+         */
+        private void checkModelMatch(String sessionId) {
+            if (modelMismatchReported.get() || cfg.model() == null || cfg.model().isBlank()) {
+                return;
+            }
+            if (sessionId == null || sessionId.isBlank()) {
+                return; // id not resolved yet — UNKNOWN, never a mismatch (fleetd #175's rule)
+            }
+            OpenCodeSessionDiscovery.ActualModel actual = discovery.actualModelForSessionId(sessionId);
+            if (actual == null) {
+                return; // UNKNOWN evidence — never a mismatch
+            }
+            String[] requestedParts = splitProviderModel(cfg.model());
+            String requestedProvider = requestedParts == null ? null : requestedParts[0];
+            String requestedId = requestedParts == null ? cfg.model() : requestedParts[1];
+            boolean idMatches = requestedId.equals(actual.id());
+            // Compare the provider ONLY when BOTH sides have one. requestedProvider == null covers
+            // a bare profile model with no "/" — the profile never asked for a specific provider.
+            // actual.provider() == null covers opencode's model JSON having an id but no providerID
+            // (a real shape parseModel accepts) — that is missing evidence, not a contradiction, and
+            // acceptance rule 4 says missing evidence is UNKNOWN, never a mismatch. Narrowing the
+            // provider comparison this way keeps the id comparison (the part that actually caught the
+            // xf bug) fully intact — a genuine id mismatch is still caught either way (fleetd #175
+            // review round 2).
+            boolean providerMatches = requestedProvider == null || actual.provider() == null
+                    || requestedProvider.equals(actual.provider());
+            if (idMatches && providerMatches) {
+                return;
+            }
+            if (!modelMismatchReported.compareAndSet(false, true)) {
+                return; // another thread already reported this exact mismatch
+            }
+            String actualDisplay = actual.provider() == null
+                    ? actual.id() : actual.provider() + "/" + actual.id();
+            log.error("opencode profile '{}' requested model '{}' but the live session is actually "
+                            + "running '{}' — opencode does not fail on an unknown -m, it silently "
+                            + "falls back to a default model, which may be a PAID credential "
+                            + "(fleetd #175); quarantining this profile's credential",
+                    cfg.profile(), cfg.model(), actualDisplay);
+            // fleetd #234: pass our OWN profile name too. This check fires from agentSessionId(),
+            // called during SessionManager.acquire() BEFORE this session is registered in
+            // sessions.roster() — a roster-only sink (Fleetd's target -> session -> profile lookup)
+            // finds nothing at this point and silently no-ops (defect 2). We already know exactly
+            // which profile to quarantine without the roster; the sink is passed it explicitly.
+            exhaustionSink.onExhausted(delegate.terminalId(),
+                    "opencode model mismatch: profile '" + cfg.profile() + "' requested '"
+                            + cfg.model() + "' but the live session is running '" + actualDisplay
+                            + "' (fleetd #175)",
+                    cfg.profile());
         }
 
         @Override

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

@@ -1,9 +1,12 @@
 package dev.ltms.fleet.member;
 
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 import org.sqlite.SQLiteConfig;
 
+import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.sql.Connection;
@@ -29,11 +32,19 @@ import java.util.concurrent.atomic.AtomicBoolean;
  * here, so a layout change, or a switch to the HTTP server, changes exactly one class and nothing
  * in {@link OpenCodeLauncher}.
  *
- * <p>The determinism that makes this useful is structural, not a guess: every fleetd worker runs
- * in its own unique git worktree, so the row's {@code directory} (its project root) equals the
- * worker's cwd identifies <em>its</em> session unambiguously. We match on {@code directory} rather
- * than diffing {@code opencode session list} before/after — that races under concurrent spawns, and
- * the CLI listing does not even show the directory.
+ * <p><strong>The {@code directory} match is a heuristic, not an identity — fleetd #234.</strong> A
+ * worktree is opt-in: {@code fleet_spawn} only provisions one when the caller passes {@code
+ * worktree:}; the default spawn inherits the lead's own cwd, which every other worker spawned the
+ * same way (and every past session ever run there) shares. {@code directory} therefore does
+ * <em>not</em> identify a session unambiguously in general — only in the special case of a fresh,
+ * unique worktree does "most recently updated row for this directory" reliably mean "this worker's
+ * own row." {@link #sessionIdForDirectory} still has to use this heuristic (the id has to come from
+ * somewhere, and nothing else is available at this layer — see that method's javadoc), but a caller
+ * that already holds a resolved id must never re-derive evidence about that same session via
+ * {@code directory} again; see {@link #actualModelForSessionId}, which looks up by {@code id}
+ * instead for exactly this reason. We match on {@code directory} rather than diffing
+ * {@code opencode session list} before/after — that races under concurrent spawns, and the CLI
+ * listing does not even show the directory.
  *
  * <p>All reads are best-effort and never throw: a missing or unreadable database, a query that
  * fails, or a directory with no row yet all yield {@code null}, and the caller (the session
@@ -44,6 +55,18 @@ import java.util.concurrent.atomic.AtomicBoolean;
 final class OpenCodeSessionDiscovery {
 
     private static final Logger log = LoggerFactory.getLogger(OpenCodeSessionDiscovery.class);
+    private static final ObjectMapper MAPPER = new ObjectMapper();
+
+    /**
+     * The model opencode actually ran a session on, parsed from the {@code session.model} JSON
+     * column (fleetd #175). {@code id} is never null/blank on a non-null {@code ActualModel} —
+     * {@link #parseModel} returns {@code null} instead when {@code id} cannot be determined, so a
+     * caller only ever sees a fully-known record or {@code null} (UNKNOWN). {@code provider} may
+     * still be {@code null} on its own when the profile that requested the session named no
+     * provider prefix, or opencode's JSON omitted {@code providerID}.
+     */
+    record ActualModel(String provider, String id) {
+    }
 
     private final Path storageRoot;   // e.g. ~/.local/share/opencode (injectable for tests)
     private final Path databasePath;
@@ -74,8 +97,16 @@ final class OpenCodeSessionDiscovery {
     /**
      * The opencode session id whose row references {@code directory} (the worker's cwd), or
      * {@code null} when no row matches yet. When several rows share the directory — e.g. repeated
-     * spawns into the same worktree — the row with the highest {@code time_updated} wins: it is
-     * the session the pane most likely corresponds to.
+     * spawns into the same worktree, OR several workers sharing one cwd because none of them was
+     * given a worktree (fleetd #234) — the row with the highest {@code time_updated} wins: it is
+     * the session the pane most likely corresponds to. That "most likely" is a real caveat, not a
+     * formality: when the directory is shared, this can and does pick another session's row (see
+     * the class javadoc). This is the one place fleetd resolves an opencode session id at all —
+     * nothing else is available at this layer to disambiguate further (no {@code opencode session
+     * list} entry names the directory, and diffing before/after races under concurrent spawns) — so
+     * the heuristic stays here unchanged. What must never happen is a SECOND, independent piece of
+     * evidence about the same session being re-derived via {@code directory} once an id has already
+     * come out of this method; see {@link #actualModelForSessionId}.
      *
      * <p>Never throws: a missing {@code opencode.db}, a locked/unreadable database, a query
      * failure, or a directory that has not been persisted yet all resolve to {@code null} rather
@@ -117,4 +148,85 @@ final class OpenCodeSessionDiscovery {
                 storageRoot, directory);
         return null;
     }
+
+    /**
+     * The model opencode actually ran the session {@code sessionId} on (fleetd #175/#234), read by
+     * primary-key lookup — the ONE row that id names, and no other. Deliberately keyed on
+     * {@code id} rather than {@code directory}: two independent {@code WHERE directory = ? ORDER BY
+     * time_updated DESC LIMIT 1} queries (one for the id, one for the model) can each pick a
+     * DIFFERENT row once more than one session shares a directory (fleetd #234 — the default
+     * no-worktree spawn shares the lead's cwd with every other worker and every past session ever
+     * run there), silently comparing a profile's requested model against a session that is not even
+     * the one whose id was returned. Keying on {@code id} instead makes that impossible: the model
+     * read back is always the SAME session {@link #sessionIdForDirectory} (or a cached copy of its
+     * answer) already resolved.
+     *
+     * <p>Kept as its own query and its own connection, independent from {@link
+     * #sessionIdForDirectory}: a database whose schema predates the {@code model} column (or any
+     * other read failure on this column alone) can never take id resolution down with it. That
+     * would be a regression of the id-resolution feature #209 shipped; this method degrades on its
+     * own.
+     *
+     * <p>Never throws, and every failure mode — no matching row, a missing/unreadable database, a
+     * missing {@code model} column, a null/blank {@code model} value, or JSON that does not parse
+     * into {@code {"id": "...", "providerID": "..."}} with a non-blank {@code id} — resolves to
+     * {@code null}. That is UNKNOWN evidence, not a mismatch signal: the caller must never
+     * quarantine a profile on the strength of a {@code null} here. A blank/null {@code sessionId}
+     * (the id is not resolved yet) is UNKNOWN too, for the same reason — never call this with one.
+     *
+     * @param sessionId the session id already resolved by {@link #sessionIdForDirectory} for this
+     *                   spawn — never re-derived from {@code directory} here
+     * @return the actual model, or {@code null} when unknown
+     */
+    ActualModel actualModelForSessionId(String sessionId) {
+        if (sessionId == null || sessionId.isBlank()) {
+            return null;
+        }
+        if (!Files.isRegularFile(databasePath)) {
+            // sessionIdForDirectory already WARNs once (shared warnedMissingDatabase) for this
+            // exact condition — do not double-log it here.
+            return null;
+        }
+        String sql = "SELECT model FROM session WHERE id = ?";
+        try (Connection connection = openReadOnly();
+             PreparedStatement statement = connection.prepareStatement(sql)) {
+            statement.setString(1, sessionId);
+            try (ResultSet rows = statement.executeQuery()) {
+                if (rows.next()) {
+                    return parseModel(rows.getString("model"));
+                }
+            }
+        } catch (SQLException e) {
+            // Locked/corrupt database, or a `model` column this schema version does not have —
+            // never fatal, and never a mismatch signal. See the class doc above.
+            log.debug("opencode session model unreadable at {}: {}", databasePath, e.toString());
+            return null;
+        }
+        return null;
+    }
+
+    /**
+     * Parse opencode's {@code model} column — {@code {"id":"...","providerID":"..."}} — into an
+     * {@link ActualModel}, or {@code null} when {@code json} is null/blank, is not valid JSON, or
+     * parses without a non-blank {@code id}. {@code providerID} may be absent; that alone does not
+     * make the record unknown, since a caller comparing against a profile with no provider prefix
+     * never looks at it.
+     */
+    private static ActualModel parseModel(String json) {
+        if (json == null || json.isBlank()) {
+            return null;
+        }
+        try {
+            JsonNode node = MAPPER.readTree(json);
+            String id = node.path("id").asText(null);
+            if (id == null || id.isBlank()) {
+                return null;
+            }
+            String provider = node.path("providerID").asText(null);
+            return new ActualModel(provider, id);
+        } catch (IOException e) {
+            log.debug("opencode session model JSON unparseable: {}", e.toString());
+            return null;
+        }
+    }
 }

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

@@ -13,6 +13,7 @@ import java.nio.charset.StandardCharsets;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.nio.file.StandardCopyOption;
+import java.nio.file.attribute.PosixFilePermissions;
 import java.security.SecureRandom;
 import java.util.ArrayList;
 import java.util.HashSet;
@@ -161,6 +162,7 @@ public final class GitWorktrees implements Worktrees {
         } catch (IOException e) {
             throw new WorktreeException("cannot create worktree root " + root + ": " + e.getMessage(), e);
         }
+        shareRootWithGroup(root);
         String wt = path.toAbsolutePath().toString();
         log.info("adding worktree branch={} path={} base={}", branch, wt, base);
         removeUserInfoFromHttpsOrigin(repoRoot);
@@ -712,6 +714,54 @@ public final class GitWorktrees implements Worktrees {
                 group, repoRoot, worktreePath, touched);
     }
 
+    /**
+     * fleetd #224: make {@code worktreeRoot} ITSELF group-traversable — established once, here,
+     * where the root is created, never at a use site. {@link #shareWithGroup} shares each worktree
+     * (and the repo's common git dir) with {@link #group}, but never the PARENT directory that
+     * contains every worktree — and under {@code memberHerdrSocket:} the member pane runs as a
+     * different OS user, which needs the execute bit on every ancestor directory to reach anything
+     * underneath, no matter how carefully each child is shared. Without this, a member cannot read
+     * the ephemeral {@code opencode.json} #219 places under this root, cannot reach its own
+     * worktree, and cannot read #213's ZDOTDIR scrub when placed here either.
+     *
+     * <p>No-op — no process spawned — when {@link #group} is null/blank, so behaviour with
+     * {@code worktreeGroup:} unset (today's only live mode) is unchanged. Only {@code root} itself
+     * is touched (non-recursive): each child underneath is shared individually, either by
+     * {@link #shareWithGroup} for a worktree or by the launcher that generates it (fleetd #213/#219)
+     * for a scrub/config directory — sharing this level again would just duplicate that policy in
+     * the wrong layer.
+     *
+     * <p>Fails loudly, naming {@code root}, its mode at the time of the attempt, and {@link #group}:
+     * a member that starts and then cannot see its own checkout is worse than a refused spawn, since
+     * nothing about that failure mode points at a directory's permission bits.
+     */
+    private void shareRootWithGroup(Path root) {
+        if (group == null) {
+            return;
+        }
+        String mode = currentPosixMode(root);
+        try {
+            shareGroupRunner.apply(new String[]{"chgrp", group, root.toString()});
+            shareGroupRunner.apply(new String[]{"chmod", "g+x", root.toString()});
+        } catch (WorktreeException e) {
+            throw new WorktreeException("cannot make worktree root " + root + " (mode " + mode
+                    + ") group-traversable for group '" + group + "': " + e.getMessage()
+                    + " — the group must exist, and the fleetd operator (" + System.getProperty("user.name")
+                    + ") must be a member of it", e);
+        }
+        log.info("worktreeGroup={} made worktree root {} group-traversable (was mode {})", group, root, mode);
+    }
+
+    /** {@code root}'s current POSIX permission string, or {@code "unknown"} on a filesystem that does
+     *  not support POSIX permissions — used only to name the mode in a refusal message. */
+    private static String currentPosixMode(Path root) {
+        try {
+            return PosixFilePermissions.toString(Files.getPosixFilePermissions(root));
+        } catch (IOException | UnsupportedOperationException e) {
+            return "unknown";
+        }
+    }
+
     /**
      * The repo's <em>common</em> git directory as an absolute path — where {@code objects},
      * {@code refs} and {@code worktrees} actually live. {@code git rev-parse --git-common-dir}

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

@@ -183,46 +183,51 @@ public final class SessionManager implements TurnListener {
                                  String sessionName, String resumeSessionId) {
         MemberRole memberRole = (role == null) ? MemberRole.DEV : role;
         requireResumeCapability(profile, resumeSessionId);
+        // CB-619 / fleetd #123: an explicit profile bypasses placement (CompositePeerLauncher only
+        // constrains an UNQUALIFIED spawn to the role's pool), so it is the one path that can ask
+        // for a role with no slot to bind it to. Refuse before anything spawns. A blank profile is
+        // left to placement, which already restricts an unqualified spawn to the role's pool.
+        if (profile != null && !profile.isBlank()) {
+            memberLifecycle.requireSlotFor(memberRole, profile);
+        }
+        MemberLifecycle.SlotReservation reservation = memberLifecycle.reserve(memberRole, profile);
+        String launchProfile = reservation == null ? profile : reservation.profile();
         if (wt == null) {
             // CB-557: the role must ride on the SpawnRequest, not stay a local. The launcher needs it
             // to pick the profile out of that role's pool and to label the tab; a role kept only on
             // the MemberSession is recorded after the spawn it was supposed to steer.
-            SpawnRequest req = new SpawnRequest(profile, requestedCwd, callerCwd, sessionName, resumeSessionId, memberRole);
+            SpawnRequest req = new SpawnRequest(launchProfile, requestedCwd, callerCwd, sessionName, resumeSessionId, memberRole);
             PeerHandle handle;
+            boolean bound = false;
             try {
                 handle = launcher.spawn(req);
+                String resolvedProfile = resolveProfile(handle, launchProfile);
+                String cwd = launcher.effectiveCwd(new SpawnRequest(resolvedProfile, requestedCwd, callerCwd));
+                long now = nowNanos.getAsLong();
+                MemberRole actualRole = acquired(memberRole, resolvedProfile, handle.terminalId(), reservation);
+                bound = reservation == null || actualRole == MemberRole.ARCHITECT;
+                MemberSession session = new MemberSession(
+                        handle.id(), handle.terminalId(), resolvedProfile, actualRole, cwd, ownerTerminal, now, now, 0,
+                        MemberSession.State.SPAWNING, null, null, handle.charterReceipt(), handle.agentSessionId());
+                registry.put(handle.id(), session);
+                handles.put(handle.id(), handle);
+                log.debug("acquired session id={} terminal={} profile={} owner={}",
+                        handle.id(), handle.terminalId(), session.profile(), session.ownerTerminal());
+                notifyAcquired(session.terminalId());
+                return session;
             } catch (RuntimeException e) {
-                log.warn("spawn failed for profile={} role={}: {}", profile, memberRole, e.getMessage());
+                if (!bound) memberLifecycle.release(reservation);
+                log.warn("spawn failed for profile={} role={}: {}", launchProfile, memberRole, e.getMessage());
                 throw e;
             }
-            String resolvedProfile = resolveProfile(handle, profile);
-            String cwd = launcher.effectiveCwd(new SpawnRequest(resolvedProfile, requestedCwd, callerCwd));
-            long now = nowNanos.getAsLong();
-            MemberSession session = new MemberSession(
-                    handle.id(),
-                    handle.terminalId(),
-                    resolvedProfile,
-                    memberRole,
-                    cwd,
-                    ownerTerminal,
-                    now,
-                    now,
-                    0,
-                    MemberSession.State.SPAWNING,
-                    null,
-                    null,
-                    handle.charterReceipt(),
-                    handle.agentSessionId());
-            registry.put(handle.id(), session);
-            handles.put(handle.id(), handle);
-            memberLifecycle.acquired(session.role(), session.profile(), session.terminalId());
-            log.debug("acquired session id={} terminal={} profile={} owner={}",
-                    handle.id(), handle.terminalId(), session.profile(), session.ownerTerminal());
-            notifyAcquired(session.terminalId());
-            return session;
         }
-        return acquireWithWorktree(profile, memberRole, requestedCwd, callerCwd, ownerTerminal, wt,
-                sessionName, resumeSessionId);
+        try {
+            return acquireWithWorktree(launchProfile, memberRole, requestedCwd, callerCwd, ownerTerminal, wt,
+                    sessionName, resumeSessionId, reservation);
+        } catch (RuntimeException e) {
+            memberLifecycle.release(reservation);
+            throw e;
+        }
     }
 
     /**
@@ -468,7 +473,8 @@ public final class SessionManager implements TurnListener {
 
     private MemberSession acquireWithWorktree(String profile, MemberRole memberRole, String requestedCwd, String callerCwd,
                                               String ownerTerminal, WorktreeRequest wt,
-                                              String sessionName, String resumeSessionId) {
+                                               String sessionName, String resumeSessionId,
+                                               MemberLifecycle.SlotReservation reservation) {
         String preResolvedProfile = (profile == null || profile.isBlank())
                 ? launcher.defaultProfile() : profile;
         // CB-507: resolve through the launcher's CB-112 chain (requested → profile cwd → caller →
@@ -509,11 +515,14 @@ public final class SessionManager implements TurnListener {
         String resolvedProfile = resolveProfile(handle, profile);
         String cwd = launcher.effectiveCwd(new SpawnRequest(resolvedProfile, path, callerCwd));
         long now = nowNanos.getAsLong();
+        // CB-619: see the no-worktree path above — bind before recording, and store the returned
+        // actual role, so this session's role is never a lie about what it actually holds.
+        MemberRole actualRole = acquired(memberRole, resolvedProfile, handle.terminalId(), reservation);
         MemberSession session = new MemberSession(
                 handle.id(),
                 handle.terminalId(),
                 resolvedProfile,
-                memberRole,
+                actualRole,
                 cwd,
                 ownerTerminal,
                 now,
@@ -526,13 +535,24 @@ public final class SessionManager implements TurnListener {
                 handle.agentSessionId());
         registry.put(handle.id(), session);
         handles.put(handle.id(), handle);
-        memberLifecycle.acquired(session.role(), session.profile(), session.terminalId());
         log.debug("acquired worktree session id={} terminal={} profile={} branch={} path={}",
                 handle.id(), handle.terminalId(), session.profile(), session.branch(), session.worktree());
         notifyAcquired(session.terminalId());
         return session;
     }
 
+    private MemberRole acquired(MemberRole role, String profile, String terminal,
+                                MemberLifecycle.SlotReservation reservation) {
+        if (reservation == null) {
+            return memberLifecycle.acquired(role, profile, terminal);
+        }
+        if (memberLifecycle.bind(reservation, terminal)) {
+            return role;
+        }
+        memberLifecycle.release(reservation);
+        return memberLifecycle.acquired(role, profile, terminal);
+    }
+
     private String slug(String raw) {
         return raw == null ? "ticket" : raw.toLowerCase().replaceAll("[^a-z0-9]+", "-").replaceAll("^-+|-+$", "");
     }

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

@@ -44,10 +44,13 @@ public final class FakeHerdr implements HerdrClient {
     private String agentSendErrorCode = null;
     private boolean noPanes = false;
     private volatile String agentStatus = "idle"; // steady-state agent.get status
+    private volatile String agentType = "claude"; // detected agent kind on agent.get; null = undetected
     private volatile String readText = "worker transcript tail"; // canned agent.read output
     private int pinnedStarts = 0;   // how many upcoming agent.start calls report a fixed pane
     private String pinnedStartTerminal;
     private String pinnedStartPane;
+    private volatile int agentGetOkCalls = Integer.MAX_VALUE; // how many agent.get calls succeed first
+    private volatile String agentGetFailCode = null; // error code every agent.get call after that reports
 
     public FakeHerdr healthy(boolean h) {
         this.healthy = h;
@@ -103,6 +106,27 @@ public final class FakeHerdr implements HerdrClient {
         return this;
     }
 
+    /**
+     * Set the detected agent kind ({@code "agent"} field) that {@code agent.get} reports —
+     * {@code null} models herdr not (or no longer) detecting a supported backend in the pane, e.g.
+     * a bare shell prompt (fleetd #176 fix 2 corroboration test).
+     */
+    public FakeHerdr agentType(String type) {
+        this.agentType = type;
+        return this;
+    }
+
+    /**
+     * Make {@code agent.get} succeed normally for its first {@code okCalls} invocations, then fail
+     * every call after that with {@code code} — fleetd #176 fix 1's "backend exited mid-wait"
+     * fixture. {@code okCalls == 0} fails from the very first call.
+     */
+    public FakeHerdr agentGetFailsWithAfter(int okCalls, String code) {
+        this.agentGetOkCalls = okCalls;
+        this.agentGetFailCode = code;
+        return this;
+    }
+
     /** The text {@code agent.read} returns (the CB-106 completion scrape). */
     public FakeHerdr readText(String text) {
         this.readText = text;
@@ -208,10 +232,21 @@ public final class FakeHerdr implements HerdrClient {
                     }
                     yield mapper.readTree("{\"type\":\"ok\"}");
                 }
-                case "agent.get" -> mapper.readTree(("""
-                        {"type":"agent_info","agent":{"terminal_id":"term_a","agent":"claude",
+                case "agent.get" -> {
+                    if (agentGetFailCode != null) {
+                        long getCalls = calls.stream().filter(c -> c.method().equals("agent.get")).count();
+                        if (getCalls > agentGetOkCalls) {
+                            throw new HerdrException(
+                                    "herdr error [" + agentGetFailCode + "]: agent target not found",
+                                    agentGetFailCode, null);
+                        }
+                    }
+                    String agentField = agentType == null ? "null" : "\"" + agentType + "\"";
+                    yield mapper.readTree(("""
+                        {"type":"agent_info","agent":{"terminal_id":"term_a","agent":%s,
                           "agent_status":"%s","workspace_id":"w2","tab_id":"w2:t7","pane_id":"w2:p7"}}""")
-                        .formatted(agentStatus));
+                            .formatted(agentField, agentStatus));
+                }
                 case "agent.read" -> mapper.readTree(mapper.writeValueAsString(
                         java.util.Map.of("type", "agent_read", "read", java.util.Map.of("text", readText))));
                 case "agent.start" -> {

fleetd/src/test/java/dev/ltms/fleet/inject/CompletionResolverTest.java

@@ -499,7 +499,7 @@ class CompletionResolverTest {
         Rendezvous rendezvous = new Rendezvous();
         ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
         java.util.List<String> notified = new java.util.ArrayList<>();
-        ExhaustionSink sink = (target, reason) -> notified.add(target + ": " + reason);
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
         CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
 
         var waiter = rendezvous.open("term_a");
@@ -520,7 +520,7 @@ class CompletionResolverTest {
         Rendezvous rendezvous = new Rendezvous();
         ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
         java.util.List<String> notified = new java.util.ArrayList<>();
-        ExhaustionSink sink = (target, reason) -> notified.add(target + ": " + reason);
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
         CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
 
         var waiter = rendezvous.open("term_a");
@@ -648,7 +648,7 @@ class CompletionResolverTest {
         Rendezvous rendezvous = new Rendezvous();
         ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
         java.util.List<String> notified = new java.util.ArrayList<>();
-        ExhaustionSink sink = (target, reason) -> notified.add(target + ": " + reason);
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
         CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
 
         var waiter = rendezvous.open("term_a");
@@ -701,7 +701,7 @@ class CompletionResolverTest {
         Rendezvous rendezvous = new Rendezvous();
         ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
         java.util.List<String> notified = new java.util.ArrayList<>();
-        ExhaustionSink sink = (target, reason) -> notified.add(target + ": " + reason);
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
         CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
 
         var waiter = rendezvous.open("term_a");
@@ -728,7 +728,7 @@ class CompletionResolverTest {
         Rendezvous rendezvous = new Rendezvous();
         ExhaustedPatternLookup patterns = target -> Pattern.compile("usage limit has been reached");
         java.util.List<String> notified = new java.util.ArrayList<>();
-        ExhaustionSink sink = (target, reason) -> notified.add(target + ": " + reason);
+        ExhaustionSink sink = (target, reason, profile) -> notified.add(target + ": " + reason);
         CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
 
         var waiter = rendezvous.open("term_a");

fleetd/src/test/java/dev/ltms/fleet/inject/ExhaustionSinkForwardingHazardTest.java

@@ -0,0 +1,41 @@
+package dev.ltms.fleet.inject;
+
+import org.junit.jupiter.api.Test;
+
+import java.util.concurrent.atomic.AtomicReference;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+
+/**
+ * fleetd #234. Round 3: calls the REAL production factory, {@link ExhaustionSink#forwardingTo},
+ * rather than rebuilding its shape locally — a round-2 version of this test built its own copy of
+ * the forwarder inline, so mutating {@code Fleetd.java}'s actual forwarder back into a lambda left
+ * that copy untouched and the test kept passing while production had regressed. Calling the shared
+ * factory here means the same object under test IS the object {@code Fleetd.java} builds via
+ * {@code ExhaustionSink.forwardingTo(exhaustionSinkRef::get)}.
+ *
+ * <p>Round 4: the 3-arg overload is now {@link ExhaustionSink}'s single abstract method, so {@code
+ * forwardingTo} itself is a one-line lambda and every sink below can safely be one too — there is
+ * no 2-arg overload left for any of them to silently bind to instead. This test still catches a
+ * regression inside {@code forwardingTo}'s body (e.g. one that calls the 2-arg default and drops
+ * the hint that way) because it still goes through the shared factory rather than a rebuilt copy.
+ */
+class ExhaustionSinkForwardingHazardTest {
+
+    @Test
+    void forwardingToDeliversTheProfileHintToWhateverSinkTheSupplierCurrentlyReturns() {
+        AtomicReference<String> hintSeenByRealSink = new AtomicReference<>("NEVER CALLED");
+        ExhaustionSink real = (target, reason, profileHint) -> hintSeenByRealSink.set(profileHint);
+
+        // Same shape Fleetd.java uses: a reference that starts at none() and is repointed later.
+        AtomicReference<ExhaustionSink> ref = new AtomicReference<>(ExhaustionSink.none());
+        ExhaustionSink forwarding = ExhaustionSink.forwardingTo(ref::get);
+        ref.set(real);
+
+        forwarding.onExhausted("term_x", "model mismatch", "gx");
+
+        assertEquals("gx", hintSeenByRealSink.get(),
+                "ExhaustionSink.forwardingTo must deliver the profile hint to the sink the supplier "
+                        + "currently returns — the exact object Fleetd.java's forwarder is built from");
+    }
+}

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

@@ -1,10 +1,13 @@
 package dev.ltms.fleet.mcp;
 
+import dev.ltms.fleet.auth.CallerResolver;
+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;
 import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.PaneLocator;
 import dev.ltms.fleet.herdr.WorkspaceControl;
 import dev.ltms.fleet.inject.Injector;
 import dev.ltms.fleet.msg.MessageService;
@@ -1024,6 +1027,87 @@ class FleetMcpTest {
         assertTrue(textOf(res).contains("architect, dev, reviewer"), textOf(res));
     }
 
+    // ── CB-619 / fleetd #123: a spawn asking for a role its profile has no slot for must be
+    //    refused, never silently demoted with the roster still lying about it ───────────────────
+
+    /** Two profiles on one launcher, so a role's pool can name one and exclude the other. */
+    private static ClaudeCodeLauncher architectCapableLauncher(FakeHerdr h) {
+        FleetConfig.Profile opus = new FleetConfig.Profile(
+                "opus", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN", null,
+                "tab", "fleetd-workers", "worker: {profile} #{n}", null, null, null);
+        FleetConfig.Profile sonnet = new FleetConfig.Profile(
+                "sonnet", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN", null,
+                "tab", "fleetd-workers", "worker: {profile} #{n}", null, null, null);
+        return new ClaudeCodeLauncher(new AgentControl(h), new WorkspaceControl(h),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of("opus", opus, "sonnet", sonnet),
+                "sonnet", _ -> "tok");
+    }
+
+    /** {@code fleet.architects} carries only {@code opus} — {@code sonnet} has no matching slot. */
+    private static MemberRegistry architectRegistry() {
+        return new MemberRegistry(new FleetConfig.Fleet(Map.of(),
+                Map.of("opus", new FleetConfig.Slot("opus")), Map.of(), Map.of(), null));
+    }
+
+    /**
+     * The literal defect (fleetd #123): {@code role=architect, profile=sonnet}, where
+     * {@code fleet.architects} carries only {@code opus}. Drives the real path —
+     * {@link FleetMcp#spawn} calls {@link SessionManager#acquire}, which must refuse before ever
+     * reaching the real {@link ClaudeCodeLauncher} — never {@link MemberRegistry#bind} called
+     * directly, which would walk around the gate under test.
+     */
+    @Test
+    void spawnRefusesAnArchitectWithNoMatchingSlotAndNeverTouchesTheLauncher() {
+        FakeHerdr h = new FakeHerdr();
+        SessionManager sessions = new SessionManager(architectCapableLauncher(h));
+        sessions.setMemberLifecycle(architectRegistry());
+
+        McpSchema.CallToolResult res = FleetMcp.spawn(sessions, "sonnet", "architect",
+                null, null, null, null, null, null);
+
+        assertEquals(Boolean.TRUE, res.isError(), textOf(res));
+        String msg = textOf(res);
+        assertTrue(msg.contains("architect"), "names the role asked for: " + msg);
+        assertTrue(msg.contains("sonnet"), "names the profile: " + msg);
+        assertTrue(msg.contains("opus"), "names the pool that does carry the role: " + msg);
+        assertTrue(sessions.roster().isEmpty(), "a refused spawn must register no session");
+        assertTrue(h.calls.stream().noneMatch(c -> c.method().equals("agent.start")),
+                "a refused spawn must never reach the launcher — no process should ever start");
+    }
+
+    /**
+     * Positive control / parity check: when the profile DOES carry a slot, the spawn succeeds, and
+     * {@code GET /members} ({@link FleetMcp#listFleet}) and {@code fleet_whoami}
+     * ({@link FleetMcp#whoami}) — resolved through the SAME live {@link MemberRegistry} binding via
+     * a real {@link CallerResolver}, exactly as the daemon resolves a real MCP caller — must never
+     * disagree about this one live member's role.
+     */
+    @Test
+    void rosterAndWhoamiAgreeOnceTheArchitectSlotBinds() {
+        FakeHerdr h = new FakeHerdr();
+        // Pin the spawn onto the one pane FakeHerdr's canned pane.process_info maps to WORKER_PID,
+        // so a CallerResolver can resolve THIS session's own terminal, not a fixture double.
+        h.pinNextStarts(1, "term_a", "w2:p7");
+        SessionManager sessions = new SessionManager(architectCapableLauncher(h));
+        MemberRegistry members = architectRegistry();
+        sessions.setMemberLifecycle(members);
+
+        McpSchema.CallToolResult spawnRes = FleetMcp.spawn(sessions, "opus", "architect",
+                null, null, null, null, null, null);
+        assertNotEquals(Boolean.TRUE, spawnRes.isError(), textOf(spawnRes));
+        assertTrue(textOf(spawnRes).contains("\"role\":\"architect\""), textOf(spawnRes));
+
+        String roster = textOf(FleetMcp.listFleet(architectCapableLauncher(h), sessions, Map.of(), ""));
+        assertTrue(roster.contains("\"role\":\"architect\""), "GET /members: " + roster);
+
+        ConnectionIdentity identity = new ConnectionIdentity(new PaneLocator(h), _ -> FakeHerdr.WORKER_PID);
+        CallerResolver resolver = CallerResolver.withLeadsAndMembers(identity, false, null, Map::of, members);
+        Principal caller = resolver.resolve("127.0.0.1", 42, null);
+        assertTrue(caller.isArchitect(), "fleet_whoami's own resolver must agree the terminal is bound");
+        String whoami = textOf(FleetMcp.whoami(caller, sessions));
+        assertTrue(whoami.contains("\"role\":\"architect\""), "fleet_whoami: " + whoami);
+    }
+
     // ── CB-584: fleet_spawn accepts sessionName/resumeSessionId; roster shows agentSessionId ──
 
     @Test

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

@@ -20,8 +20,10 @@ import org.junit.jupiter.api.Test;
 import org.junit.jupiter.api.io.TempDir;
 import org.slf4j.LoggerFactory;
 
+import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
+import java.nio.file.attribute.PosixFilePermissions;
 import java.util.List;
 import java.util.Map;
 import java.util.Set;
@@ -31,6 +33,7 @@ import java.util.function.Function;
 import java.util.function.Supplier;
 
 import static org.junit.jupiter.api.Assertions.*;
+import static org.junit.jupiter.api.Assumptions.assumeTrue;
 
 /** The step-4 launch-flag injection: the bridge MCP + reply charter are appended to the argv. */
 class ClaudeCodeLauncherTest {
@@ -960,6 +963,152 @@ class ClaudeCodeLauncherTest {
         assertDoesNotThrow(() -> UUID.fromString(handle.id()));
     }
 
+    // --- fleetd #176 fix 1: fail fast when the backend process exits mid-wait -------------------
+
+    @Test
+    void spawnFailsFastWhenBackendProcessExitsMidWaitInsteadOfBurningTheTimeout() {
+        // First agent.get sees UNKNOWN (one normal tick); the second reports the pane gone, exactly
+        // what herdr answers when the backend process has already exited. The timeout is generous
+        // (60s) so a test that wrongly falls through to the old unguarded call — and therefore
+        // waits out the whole window — is unambiguously distinguishable from one that fails fast.
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("unknown");
+        herdr.agentGetFailsWithAfter(1, "pane_not_found");
+        long[] clock = {0};
+
+        ClaudeCodeLauncher svc = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")),
+                workerConfigMap("ltms-local", null), "ltms-local", _ -> null,
+                60_000, () -> clock[0], () -> clock[0] += 50);
+
+        PeerUnreachableException ex = assertThrows(
+                PeerUnreachableException.class,
+                () -> svc.spawn(new SpawnRequest(null, null, null)));
+
+        assertTrue(ex.getMessage().contains("w9:pRoot_1"),
+                "exception message references the paneId: " + ex.getMessage());
+        assertTrue(ex.getMessage().toLowerCase().contains("exited"),
+                "exception message says the backend exited, not that the pane was slow: "
+                        + ex.getMessage());
+        assertTrue(clock[0] < 60_000,
+                "the gate must not burn the rest of the 60s timeout: clock only reached " + clock[0]);
+        long getCalls = herdr.calls.stream().filter(c -> c.method().equals("agent.get")).count();
+        assertEquals(2, getCalls,
+                "exactly one normal poll then the not_found answer — no further polling after that: "
+                        + getCalls);
+        assertEquals(1, paneCloseCount(herdr, "w9:pRoot_1"),
+                "the pane is torn down (no orphan) even on the fail-fast path");
+    }
+
+    @Test
+    void spawnLetsAnUnrelatedHerdrErrorPropagateUnchanged() {
+        // Fix 1 must only special-case a "*_not_found" answer. Any other herdr failure keeps
+        // propagating as-is — this gate does not know how to recover from it.
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("unknown");
+        herdr.agentGetFailsWithAfter(0, "internal_error");
+        long[] clock = {0};
+
+        ClaudeCodeLauncher svc = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")),
+                workerConfigMap("ltms-local", null), "ltms-local", _ -> null,
+                60_000, () -> clock[0], () -> clock[0] += 50);
+
+        dev.ltms.fleet.herdr.HerdrException ex = assertThrows(
+                dev.ltms.fleet.herdr.HerdrException.class,
+                () -> svc.spawn(new SpawnRequest(null, null, null)));
+
+        assertEquals("internal_error", ex.code());
+        assertEquals(0, paneCloseCount(herdr, "w9:pRoot_1"),
+                "an error this gate does not recognize is not this gate's teardown to run");
+    }
+
+    // --- fleetd #176 fix 2: corroborated UNKNOWN refinement --------------------------------------
+
+    @Test
+    void refinedIdleIsNotAcceptedWhenAgentTypeIsNull() {
+        // The trap fix 2 must close: a pane sitting at a bare shell prompt after its backend exited
+        // still contains the same "❯" glyph StatusRefiner.classify treats as "idle at the Claude
+        // Code TUI prompt". Without the agentType corroboration this would be misread as injectable
+        // and the gate would hand back a peer that never started. herdr reports no agentType for
+        // that bare shell.
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("unknown"); // always UNKNOWN
+        herdr.agentType(null);        // no supported backend detected — could be a bare shell
+        herdr.readText("some-host:~ user$ ❯ "); // looks exactly like an idle Claude Code prompt
+        long[] clock = {0};
+
+        ClaudeCodeLauncher svc = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")),
+                workerConfigMap("ltms-local", null), "ltms-local", _ -> null,
+                500, () -> clock[0], () -> clock[0] += 50);
+
+        PeerUnreachableException ex = assertThrows(
+                PeerUnreachableException.class,
+                () -> svc.spawn(new SpawnRequest(null, null, null)));
+
+        assertTrue(clock[0] >= 500,
+                "a null agentType must not let the '❯' prompt refine to injectable — the gate has "
+                        + "to wait out the full timeout: clock only reached " + clock[0]);
+        assertEquals(1, paneCloseCount(herdr, "w9:pRoot_1"),
+                "the pane is torn down on timeout, same as any other never-injectable spawn");
+    }
+
+    @Test
+    void refinedIdleIsAcceptedWhenAgentTypeCorroboratesLiveness() {
+        // The positive case: a genuinely live claude pane that herdr misreports as UNKNOWN (CB-115)
+        // still resolves to injectable once agentType corroborates that a supported backend is
+        // detected in the same sample.
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("unknown"); // always UNKNOWN at the raw level
+        herdr.agentType("claude");    // herdr still detects a live claude backend
+        herdr.readText("? for shortcuts"); // StatusRefiner.classify's idle footer marker
+        long[] clock = {0};
+
+        ClaudeCodeLauncher svc = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")),
+                workerConfigMap("ltms-local", null), "ltms-local", _ -> null,
+                60_000, () -> clock[0], () -> clock[0] += 50);
+
+        PeerHandle handle = svc.spawn(new SpawnRequest(null, null, null));
+
+        assertNotNull(handle, "a corroborated refined-IDLE sample lets the spawn succeed");
+        assertTrue(clock[0] < 60_000,
+                "refinement must resolve well before the timeout: clock reached " + clock[0]);
+        assertEquals(0, paneCloseCount(herdr, "w9:pRoot_1"),
+                "no pane close — the peer is genuinely injectable");
+    }
+
+    @Test
+    void refinementNeverFiresWhenRawStatusIsAlreadyInjectable() {
+        // Acceptance criterion 4: refinement must trigger only on a raw UNKNOWN sample. Seed the
+        // pane content with an active-generation marker that StatusRefiner.classify would read as
+        // WORKING (never injectable) if refine() were wrongly invoked here — proving that a raw
+        // IDLE status short-circuits before refine() (and its extra agent.read call) ever runs.
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("idle"); // already injectable at the raw level
+        herdr.agentType("claude");
+        herdr.readText("esc to interrupt"); // would classify as WORKING if refine() ran anyway
+
+        long[] clock = {0};
+        ClaudeCodeLauncher gated = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")),
+                workerConfigMap("ltms-local", null), "ltms-local", _ -> null,
+                5000, () -> clock[0], () -> clock[0] += 300);
+
+        PeerHandle handle = gated.spawn(new SpawnRequest(null, null, null));
+
+        assertNotNull(handle, "an already-injectable raw status succeeds without ever refining");
+        assertFalse(herdr.called("agent.read"),
+                "refine() must never run (and so never call agent.read) when the raw status is "
+                        + "already injectable");
+    }
+
     // --- CB-511: worker environment seeding -----------------------------------------------------
 
     @Test
@@ -1751,4 +1900,184 @@ class ClaudeCodeLauncherTest {
 
         assertEquals(List.of("dev: sonnet #1", "[sonnet] dev 2"), tabLabels(herdr));
     }
+
+    // ── fleetd #222: the charter file must not land under fleetd's own java.io.tmpdir when the ──
+    // ── member pane runs as a different OS user ─────────────────────────────────────────────────
+
+    /** A profile that mounts the bridge MCP (so a reply charter is always generated). */
+    private static FleetConfig.Profile charterCfg() {
+        return new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
+                List.of("claude"), "tab", "fleetd-workers", "worker: {profile} #{n}",
+                "http://127.0.0.1:8765/mcp", null, null);
+    }
+
+    /** A config with {@code memberHerdrSocket:} set, and optionally {@code worktreeRoot:}/{@code worktreeGroup:}. */
+    private static FleetConfig configWithMemberHerdrSocket(String worktreeRoot, String worktreeGroup) {
+        return new FleetConfig(
+                null,                    // bind
+                null,                    // herdrSocket
+                "/tmp/other-user.sock",  // memberHerdrSocket
+                Map.of(),                // profiles
+                null,                    // guard
+                worktreeRoot,            // worktreeRoot
+                null,                    // lifecycle
+                null,                    // spawnReadyTimeoutMs
+                null,                    // spawnReadyPollMs
+                null,                    // broker
+                null,                    // primary
+                null,                    // fleet
+                null,                    // leadHeartbeat
+                null,                    // health
+                null,                    // placement
+                null,                    // auth
+                null,                    // configReload
+                null,                    // quarantineCooldownSeconds
+                null,                    // memberCredentials
+                null,                    // coordinator
+                worktreeGroup,           // worktreeGroup
+                null                     // memberLoginShell
+        ).withDefaults();
+    }
+
+    private static ClaudeCodeLauncher serviceWithConfig(FakeHerdr herdr, FleetConfig.Profile cfg,
+                                                          Supplier<FleetConfig> config) {
+        return new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null,
+                0, System::currentTimeMillis, () -> { }, null, null, null, config);
+    }
+
+    /**
+     * The current process's REAL primary group — resolved via {@code id -gn}, never by reading a
+     * directory's owning group (fleetd #225). Those two coincide only by accident: a directory's
+     * group is whichever group happened to own the path Maven was started from — {@code staff} in
+     * a home checkout, {@code wheel} under {@code /private/tmp} on macOS — and the fix-up this test
+     * exercises then fails for real when the operator is not a member of that borrowed group,
+     * exactly the case {@code assumeTrue(view != null, ...)} never covered (it only detects a
+     * filesystem with no POSIX groups at all, not a resolvable-but-wrong one). Skips (never fails)
+     * when {@code id} is unavailable or its primary group cannot be resolved on this host.
+     */
+    private static String currentUserGroup() {
+        String out;
+        boolean ok;
+        try {
+            Process p = new ProcessBuilder("id", "-gn").redirectErrorStream(true).start();
+            try (java.io.BufferedReader r = new java.io.BufferedReader(
+                    new java.io.InputStreamReader(p.getInputStream(), java.nio.charset.StandardCharsets.UTF_8))) {
+                out = r.lines().collect(java.util.stream.Collectors.joining("\n")).trim();
+            }
+            ok = p.waitFor(5, java.util.concurrent.TimeUnit.SECONDS) && p.exitValue() == 0 && !out.isBlank();
+        } catch (IOException e) {
+            out = null;
+            ok = false;
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            out = null;
+            ok = false;
+        }
+        assumeTrue(ok, "cannot resolve this process's real primary group via `id -gn` on this host "
+                + "— skipping a POSIX-group-dependent test rather than failing it");
+        return out;
+    }
+
+    /**
+     * fleetd #222 acceptance criterion 1: with {@code memberHerdrSocket} configured and both
+     * {@code worktreeRoot}/{@code worktreeGroup} set, the charter file lives in a fresh per-spawn
+     * directory under {@code worktreeRoot} — NEVER under {@code java.io.tmpdir} (fleetd's own 0700
+     * temp dir, unreadable by the member's different OS user) — and that directory is shared
+     * read-only with the group via the SAME mechanism (fleetd #213/#219's {@link
+     * EnvAllowListScrub#shareWithGroup}) the ZDOTDIR scrub and the opencode config directory use.
+     */
+    @Test
+    void memberHerdrSocketWithWorktreeRootAndGroupPutsCharterUnderWorktreeRootAndSharesIt(
+            @TempDir Path worktreeRoot) throws Exception {
+        String group = currentUserGroup();
+        FakeHerdr herdr = new FakeHerdr();
+        serviceWithConfig(herdr, charterCfg(),
+                () -> configWithMemberHerdrSocket(worktreeRoot.toString(), group)).spawn();
+
+        List<String> args = spawnedArgs(herdr);
+        int fileFlag = args.indexOf("--append-system-prompt-file");
+        assertTrue(fileFlag >= 0, "the charter is still mounted via file: " + args);
+        Path charterFile = Path.of(args.get(fileFlag + 1));
+        Path dir = charterFile.getParent();
+
+        // NOTE: JUnit's own @TempDir provider places worktreeRoot itself under java.io.tmpdir on this
+        // host, so "not under java.io.tmpdir" is not a meaningful assertion here (it would hold by
+        // accident of the fixture, not by anything this method does). What this fix actually promises
+        // is that the directory is created UNDER worktreeRoot specifically — never resolved from the
+        // no-argument Files.createTempFile default (java.io.tmpdir) the pre-fix code always used — so
+        // that is the assertion: the parent is exactly worktreeRoot, whatever directory JUnit gave it.
+        assertEquals(worktreeRoot.toAbsolutePath().normalize(), dir.getParent(),
+                "the generated directory's parent must be worktreeRoot, not java.io.tmpdir — got "
+                        + "parent " + dir.getParent());
+
+        assertEquals("rwxr-x---", PosixFilePermissions.toString(Files.getPosixFilePermissions(dir)),
+                "the directory must be group-traversable+readable, owner-only writable");
+        assertEquals("rw-r-----", PosixFilePermissions.toString(Files.getPosixFilePermissions(charterFile)),
+                "the charter file must be group-readable, never group-writable");
+        assertTrue(Files.readString(charterFile).contains("fleet_reply"),
+                "the charter content itself is unaffected by where it is written");
+    }
+
+    /**
+     * fleetd #222 acceptance criterion 2: with {@code memberHerdrSocket} configured but NEITHER
+     * {@code worktreeRoot} nor {@code worktreeGroup} set, the launcher must refuse the spawn rather
+     * than hand the member a {@code --append-system-prompt-file} path under {@code java.io.tmpdir}
+     * it cannot read — the member's whole turn contract would never reach it.
+     */
+    @Test
+    void memberHerdrSocketWithoutWorktreeRootOrGroupRefusesTheSpawn() {
+        FakeHerdr herdr = new FakeHerdr();
+        ClaudeCodeLauncher launcher = serviceWithConfig(herdr, charterCfg(),
+                () -> configWithMemberHerdrSocket(null, null));
+
+        IllegalStateException ex = assertThrows(IllegalStateException.class, launcher::spawn,
+                "a missing worktreeRoot/worktreeGroup must refuse the spawn, not write an unreadable charter");
+        assertTrue(ex.getMessage().contains("worktreeRoot"),
+                "the refusal must name the missing config key — got: " + ex.getMessage());
+        assertFalse(herdr.called("agent.start"),
+                "the spawn must be refused BEFORE the member is ever started — got calls: " + herdr.calls);
+    }
+
+    /**
+     * fleetd #222 acceptance criterion 2 (the other missing half): {@code worktreeRoot} set but
+     * {@code worktreeGroup} missing must ALSO refuse — either one alone is not enough to guarantee
+     * the member's OS user can read the charter file.
+     */
+    @Test
+    void memberHerdrSocketWithWorktreeRootButNoGroupRefusesTheSpawn(@TempDir Path worktreeRoot) {
+        FakeHerdr herdr = new FakeHerdr();
+        ClaudeCodeLauncher launcher = serviceWithConfig(herdr, charterCfg(),
+                () -> configWithMemberHerdrSocket(worktreeRoot.toString(), null));
+
+        IllegalStateException ex = assertThrows(IllegalStateException.class, launcher::spawn);
+        assertTrue(ex.getMessage().contains("worktreeGroup"),
+                "worktreeRoot alone is not enough — got: " + ex.getMessage());
+    }
+
+    /**
+     * fleetd #222 acceptance criterion 3: with {@code memberHerdrSocket} ABSENT — even when a live,
+     * non-null {@code config} supplier is threaded through (not merely {@code config == null}, which
+     * every other test in this file already exercises) — the charter file must still be created
+     * directly under {@code java.io.tmpdir} via the same no-directory-argument
+     * {@code Files.createTempFile} call as before this fix, byte-identical to today.
+     */
+    @Test
+    void memberHerdrSocketAbsentStaysUnderJavaIoTmpdirEvenWithALiveConfigSupplier() throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig config = new FleetConfig(null, null, null, Map.of(), null, null, null, null, null,
+                null, null, null, null, null, null, null, null, null, null, null, null, null).withDefaults();
+        serviceWithConfig(herdr, charterCfg(), () -> config).spawn();
+
+        List<String> args = spawnedArgs(herdr);
+        int fileFlag = args.indexOf("--append-system-prompt-file");
+        assertTrue(fileFlag >= 0);
+        Path charterFile = Path.of(args.get(fileFlag + 1));
+        assertTrue(charterFile.startsWith(Path.of(System.getProperty("java.io.tmpdir"))),
+                "with memberHerdrSocket absent, the charter file must still land directly under "
+                        + "java.io.tmpdir, unchanged from before this fix");
+        assertTrue(charterFile.getFileName().toString().startsWith("fleetd-role-charter-"),
+                "same file-naming scheme as before this fix (no wrapping directory): " + charterFile);
+    }
 }

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

@@ -114,6 +114,68 @@ class EnvAllowListScrubTest {
         assertNull(EnvAllowListScrub.readReport(dir));
     }
 
+    /**
+     * fleetd #224 criterion 5 (a gap the #221 reviewer flagged): {@link EnvAllowListScrub#shareWithGroup}
+     * lists one FLAT level of {@code dir} and shares every file it finds there — which does cover the
+     * OPTIONAL files a caller may or may not have written before calling it ({@code member-charter.md},
+     * {@code ide-rules.md} — both written by {@code OpenCodeLauncher#writeConfig}), but nothing
+     * pinned that down.
+     * Without this test, a future change that writes a file AFTER the sharing call, or into a
+     * subdirectory, would pass every existing test while quietly leaving that file unreadable to a
+     * different-uid member.
+     *
+     * <p>This drives {@code shareWithGroup} directly against a directory holding several flat files —
+     * not only {@code opencode.json}, but also the two optional ones named above plus a third,
+     * unrelated file, so the assertion is "every flat file", not "the two files someone thought of".
+     */
+    @Test
+    void shareWithGroupCoversEveryFlatFileIncludingTheOptionalOnes(@TempDir Path dir) throws Exception {
+        String group = currentUserGroup();
+        Files.writeString(dir.resolve("opencode.json"), "{}\n");
+        Files.writeString(dir.resolve("member-charter.md"), "# charter\n");
+        Files.writeString(dir.resolve("ide-rules.md"), "# ide rules\n");
+        Files.writeString(dir.resolve("another-flat-file.txt"), "unrelated\n");
+
+        EnvAllowListScrub.shareWithGroup(dir, group);
+
+        assertEquals("rwxr-x---", java.nio.file.attribute.PosixFilePermissions.toString(
+                        Files.getPosixFilePermissions(dir)),
+                "the directory itself must be group-traversable+readable, owner-only writable");
+        for (String name : List.of("opencode.json", "member-charter.md", "ide-rules.md", "another-flat-file.txt")) {
+            Path file = dir.resolve(name);
+            assertEquals("rw-r-----", java.nio.file.attribute.PosixFilePermissions.toString(
+                            Files.getPosixFilePermissions(file)),
+                    name + " must be group-readable, never group-writable");
+        }
+    }
+
+    /**
+     * The current process's REAL primary group — resolved via {@code id -gn}, never by reading a
+     * directory's owning group (fleetd #225: that reads wherever Maven happened to be started from,
+     * not the process's own group, and the two diverge outside a home checkout). Skips (never fails)
+     * when {@code id} is unavailable or its primary group cannot be resolved on this host.
+     */
+    private static String currentUserGroup() {
+        String out;
+        boolean ok;
+        try {
+            Process p = new ProcessBuilder("id", "-gn").redirectErrorStream(true).start();
+            String raw = new String(p.getInputStream().readAllBytes(), StandardCharsets.UTF_8).trim();
+            out = raw;
+            ok = p.waitFor(5, java.util.concurrent.TimeUnit.SECONDS) && p.exitValue() == 0 && !raw.isBlank();
+        } catch (IOException e) {
+            out = null;
+            ok = false;
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            out = null;
+            ok = false;
+        }
+        assumeTrue(ok, "cannot resolve this process's real primary group via `id -gn` on this host "
+                + "— skipping a POSIX-group-dependent test rather than failing it");
+        return out;
+    }
+
     /**
      * The same equality, for a shell that is INTERACTIVE but NOT a login shell — the shape herdr
      * opens on Linux.

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

@@ -18,7 +18,6 @@ import org.slf4j.LoggerFactory;
 import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
-import java.nio.file.attribute.PosixFileAttributeView;
 import java.util.HashMap;
 import java.util.List;
 import java.util.Map;
@@ -513,11 +512,37 @@ class HerdrPeerLauncherAllowListWiringTest {
                         + "java.io.tmpdir, unchanged from before this fix: " + dir);
     }
 
-    /** The current process's own primary group — resolvable on whatever host runs this test. */
-    private static String currentUserGroup() throws IOException {
-        PosixFileAttributeView view = Files.getFileAttributeView(Path.of("."), PosixFileAttributeView.class);
-        assumeTrue(view != null, "this host's filesystem does not support POSIX group ownership");
-        return view.readAttributes().group().getName();
+    /**
+     * The current process's REAL primary group — resolved via {@code id -gn}, never by reading a
+     * directory's owning group (fleetd #225). Those two coincide only by accident: a directory's
+     * group is whichever group happened to own the path Maven was started from — {@code staff} in
+     * a home checkout, {@code wheel} under {@code /private/tmp} on macOS — and the fix-up this test
+     * exercises then fails for real when the operator is not a member of that borrowed group,
+     * exactly the case {@code assumeTrue(view != null, ...)} never covered (it only detects a
+     * filesystem with no POSIX groups at all, not a resolvable-but-wrong one). Skips (never fails)
+     * when {@code id} is unavailable or its primary group cannot be resolved on this host.
+     */
+    private static String currentUserGroup() {
+        String out;
+        boolean ok;
+        try {
+            Process p = new ProcessBuilder("id", "-gn").redirectErrorStream(true).start();
+            try (java.io.BufferedReader r = new java.io.BufferedReader(
+                    new java.io.InputStreamReader(p.getInputStream(), java.nio.charset.StandardCharsets.UTF_8))) {
+                out = r.lines().collect(java.util.stream.Collectors.joining("\n")).trim();
+            }
+            ok = p.waitFor(5, java.util.concurrent.TimeUnit.SECONDS) && p.exitValue() == 0 && !out.isBlank();
+        } catch (IOException e) {
+            out = null;
+            ok = false;
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            out = null;
+            ok = false;
+        }
+        assumeTrue(ok, "cannot resolve this process's real primary group via `id -gn` on this host "
+                + "— skipping a POSIX-group-dependent test rather than failing it");
+        return out;
     }
 
     /** Spawn once through the real launcher path, capturing every INFO+ line this class logs. */

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

@@ -10,12 +10,16 @@ import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.herdr.AgentControl;
 import dev.ltms.fleet.herdr.FakeHerdr;
 import dev.ltms.fleet.herdr.WorkspaceControl;
+import dev.ltms.fleet.inject.ExhaustionSink;
 import dev.ltms.fleet.peer.Capability;
 import dev.ltms.fleet.peer.CharterReceipt;
 import dev.ltms.fleet.peer.MemberRole;
 import dev.ltms.fleet.peer.PeerHandle;
 import dev.ltms.fleet.peer.PeerUnreachableException;
 import dev.ltms.fleet.peer.SpawnRequest;
+import dev.ltms.fleet.placement.BackendQuarantine;
+import dev.ltms.fleet.session.MemberSession;
+import dev.ltms.fleet.session.SessionManager;
 import org.junit.jupiter.api.Test;
 import org.junit.jupiter.api.io.TempDir;
 import org.slf4j.LoggerFactory;
@@ -23,13 +27,15 @@ import org.slf4j.LoggerFactory;
 import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
-import java.nio.file.attribute.PosixFileAttributeView;
 import java.nio.file.attribute.PosixFilePermissions;
+import java.util.ArrayList;
 import java.util.List;
 import java.util.Map;
+import java.util.Optional;
 import java.util.concurrent.ExecutorService;
 import java.util.concurrent.Executors;
 import java.util.concurrent.Future;
+import java.util.concurrent.TimeUnit;
 import java.util.function.Supplier;
 
 import static org.junit.jupiter.api.Assertions.*;
@@ -48,6 +54,19 @@ class OpenCodeLauncherTest {
                 null, null, gitTokenEnv, null, FleetConfig.Profile.KIND_OPENCODE);
     }
 
+    /**
+     * A profile carrying an explicit {@code credentialId} (fleetd #234 defect 2) — distinct from
+     * the profile's own name, so a test can assert on the credential precisely rather than relying
+     * on {@code effectiveCredentialId()}'s profile-name fallback.
+     */
+    private static FleetConfig.Profile opencodeCfgWithCredential(String profileName, String model,
+                                                                   String credentialId) {
+        return new FleetConfig.Profile(profileName, null, model, null, "FLEETD_WORKER_TOKEN",
+                List.of("opencode"), "tab", "fleetd-workers", "opencode: {model} #{n}", null,
+                null, List.of(), null, null, FleetConfig.Profile.KIND_OPENCODE, Map.of(), 1.0f,
+                null, false, null, credentialId, null);
+    }
+
     /** Gate-disabled launcher whose per-spawn config dirs land under an inspectable temp root. */
     private static OpenCodeLauncher service(FakeHerdr herdr, Path configRoot, FleetConfig.Profile cfg) {
         return new OpenCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
@@ -392,6 +411,46 @@ class OpenCodeLauncherTest {
         assertNotNull(ex.getMessage());
     }
 
+    /**
+     * fleetd #176 fix 2's per-adapter guard: {@code StatusRefiner.classify} is written for the
+     * Claude Code TUI only, and the spawn-readiness gate must never run it against an opencode
+     * pane. {@code agentType("opencode")} deliberately satisfies the OTHER guard (the corroborating
+     * liveness check) so it cannot be what makes this test pass — only the {@code namePrefix}
+     * check can be. If that check were ever removed, this pane's {@code ❯} content would refine
+     * straight to IDLE and the gate would report ready before the backend actually was.
+     */
+    @Test
+    void opencodePaneIsNeverRefinedEvenWhenItsContentLooksLikeAnIdleClaudePrompt(@TempDir Path root) {
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("unknown"); // always UNKNOWN
+        herdr.agentType("opencode");  // non-null — satisfies the liveness guard on its own
+        herdr.readText("some-host:~ user$ ❯ "); // content StatusRefiner.classify reads as IDLE
+        long[] clock = {0};
+        OpenCodeLauncher svc = new OpenCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                Map.of("gemini", opencodeCfg(null, null, null)), "gemini", _ -> null,
+                1000, () -> clock[0], () -> clock[0] += 50, root, root);
+
+        assertThrows(PeerUnreachableException.class,
+                () -> svc.spawn(new SpawnRequest(null, null, null)));
+
+        assertTrue(clock[0] >= 1000,
+                "an opencode pane must never refine to injectable, however its content reads — "
+                        + "the gate has to wait out the full timeout: clock only reached " + clock[0]);
+        // A blanket "agent.read is never called" does not hold here: the timeout path itself reads
+        // the pane tail (source=recent) for its own log message, on every timeout, regardless of
+        // adapter — see HerdrPeerLauncher.readPaneQuietly. So assert on the refiner's OWN probe
+        // source (StatusRefiner.PROBE_SOURCE = "detection") instead — that call happens only inside
+        // StatusRefiner.refine, so its absence proves the refiner itself was never reached for this
+        // opencode pane, not merely that its answer was discarded.
+        long detectionReads = herdr.calls.stream()
+                .filter(c -> c.method().equals("agent.read"))
+                .filter(c -> "detection".equals(((Map<?, ?>) c.params()).get("source")))
+                .count();
+        assertEquals(0, detectionReads,
+                "the refiner's own pane probe (source=detection) must never run against an "
+                        + "opencode pane — the namePrefix guard has to stop it before that call");
+    }
+
     @Test
     void spawnReturnsHandleWhenGateDisabled(@TempDir Path root) {
         FakeHerdr herdr = new FakeHerdr();
@@ -657,11 +716,37 @@ class OpenCodeLauncherTest {
                 0, System::currentTimeMillis, () -> { }, configRoot, discoveryRoot, null, null, config);
     }
 
-    /** The current process's own primary group — resolvable on whatever host runs this test. */
-    private static String currentUserGroup() throws IOException {
-        PosixFileAttributeView view = Files.getFileAttributeView(Path.of("."), PosixFileAttributeView.class);
-        assumeTrue(view != null, "this host's filesystem does not support POSIX group ownership");
-        return view.readAttributes().group().getName();
+    /**
+     * The current process's REAL primary group — resolved via {@code id -gn}, never by reading a
+     * directory's owning group (fleetd #225). Those two coincide only by accident: a directory's
+     * group is whichever group happened to own the path Maven was started from — {@code staff} in
+     * a home checkout, {@code wheel} under {@code /private/tmp} on macOS — and the fix-up this test
+     * exercises then fails for real when the operator is not a member of that borrowed group,
+     * exactly the case {@code assumeTrue(view != null, ...)} never covered (it only detects a
+     * filesystem with no POSIX groups at all, not a resolvable-but-wrong one). Skips (never fails)
+     * when {@code id} is unavailable or its primary group cannot be resolved on this host.
+     */
+    private static String currentUserGroup() {
+        String out;
+        boolean ok;
+        try {
+            Process p = new ProcessBuilder("id", "-gn").redirectErrorStream(true).start();
+            try (java.io.BufferedReader r = new java.io.BufferedReader(
+                    new java.io.InputStreamReader(p.getInputStream(), java.nio.charset.StandardCharsets.UTF_8))) {
+                out = r.lines().collect(java.util.stream.Collectors.joining("\n")).trim();
+            }
+            ok = p.waitFor(5, java.util.concurrent.TimeUnit.SECONDS) && p.exitValue() == 0 && !out.isBlank();
+        } catch (IOException e) {
+            out = null;
+            ok = false;
+        } catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            out = null;
+            ok = false;
+        }
+        assumeTrue(ok, "cannot resolve this process's real primary group via `id -gn` on this host "
+                + "— skipping a POSIX-group-dependent test rather than failing it");
+        return out;
     }
 
     /**
@@ -811,4 +896,452 @@ class OpenCodeLauncherTest {
         assertTrue(warnings.get(0).contains("memberHerdrSocket"),
                 "the WARN must name memberHerdrSocket as the reason — got: " + warnings.get(0));
     }
+
+    // --- fleetd #175: opencode silently substitutes a model on an unknown -m flag ----------------
+
+    private static OpenCodeLauncher serviceWithSink(FakeHerdr herdr, Path configRoot, Path discoveryRoot,
+                                                      FleetConfig.Profile cfg, ExhaustionSink sink) {
+        return new OpenCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null,
+                0, System::currentTimeMillis, () -> { }, configRoot, discoveryRoot,
+                null, null, null, sink);
+    }
+
+    /**
+     * fleetd #175 acceptance criterion 1: the check must fire on the REAL late-resolve path —
+     * {@code SessionManager}'s {@code get}/{@code roster} re-polling a retained {@link PeerHandle}
+     * (fleetd #209) — not on a handle built and queried directly. A handle whose row does not exist
+     * yet, then does, proves the check runs exactly where production runs it: PR #203 shipped a
+     * check that ran inside {@code spawn()}, before opencode had written the row, and every test
+     * passed anyway because none of them drove it through this path. This test would have caught
+     * that: {@code sessions.get(...)} before the row exists must show no mismatch, and the SAME
+     * call, re-driven after the row appears, must be what fires the sink.
+     */
+    @Test
+    void theRealSessionManagerLateResolvePathCatchesAModelMismatch(@TempDir Path configRoot,
+                                                                     @TempDir Path discRoot) throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        // xf's real shape (fleetd #175): weight:80, model "opencode/nemotron-3-ultra-free", no
+        // credentialId — the profile that actually escaped the fleet's accounting.
+        FleetConfig.Profile cfg = opencodeCfg("opencode/nemotron-3-ultra-free", null, null);
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(target + "|" + reason);
+        OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, sink);
+
+        SessionManager sessions = new SessionManager(launcher);
+        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+
+        // Real late-resolve path, driven BEFORE opencode has written its session row — same shape
+        // as production the instant a pane goes ready.
+        Optional<MemberSession> beforeRow = sessions.get(acquired.paneId());
+        assertTrue(beforeRow.isPresent());
+        assertNull(beforeRow.get().agentSessionId(), "no opencode row yet");
+        assertTrue(exhausted.isEmpty(), "no row yet → nothing to compare, the sink must stay silent");
+
+        // opencode writes its row late, running gpt-5.6-sol (a PAID credential) instead of the
+        // withdrawn free model the profile actually asked for — the exact fleetd #175 scenario.
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+
+        // Drive the SAME real late-resolve path again: sessions.get() -> resolveAgentSessionId ->
+        // the retained PeerHandle's agentSessionId() -> discovery -> checkModelMatch, all in one
+        // call, unmodified SessionManager code from fleetd #209.
+        Optional<MemberSession> afterRow = sessions.get(acquired.paneId());
+        assertEquals("ses_x", afterRow.get().agentSessionId(),
+                "the id itself still resolves correctly alongside the model check");
+
+        assertEquals(1, exhausted.size(), "the mismatch must fire exactly once through the real path");
+        assertTrue(exhausted.get(0).contains(cfg.model()), "reports the requested model: " + exhausted.get(0));
+        assertTrue(exhausted.get(0).contains("gpt-5.6-sol"), "reports the actual model: " + exhausted.get(0));
+    }
+
+    /** THE TRAP, row 1: a provider-prefixed request matching the DB's id AND provider is a match. */
+    @Test
+    void aProviderPrefixedModelMatchingBothIdAndProviderIsNotAMismatch(@TempDir Path configRoot,
+                                                                        @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
+
+        assertEquals("ses_x", handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(), "id and provider both match → never a mismatch: " + exhausted);
+    }
+
+    /** THE TRAP, row 2: same shape as row 1 with a different provider/id pair (the gx profile). */
+    @Test
+    void aGxProviderPrefixedModelMatchingBothIdAndProviderIsNotAMismatch(@TempDir Path configRoot,
+                                                                          @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("gx/deepseek-v4-flash", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
+
+        assertEquals("ses_x", handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(), "id and provider both match → never a mismatch: " + exhausted);
+    }
+
+    /**
+     * Incomplete evidence, not THE TRAP's provider mismatch: opencode's {@code model} JSON had an
+     * {@code id} but no {@code providerID} at all (a real shape {@code parseModel} accepts — see
+     * {@code OpenCodeSessionDiscoveryTest}). The id matches; the provider dimension is simply
+     * unknown, not contradicted. A provider-prefixed profile must NOT be quarantined on this —
+     * that would quarantine on incomplete evidence, which acceptance rule 4 forbids.
+     */
+    @Test
+    void aMissingProviderIdInTheEvidenceIsUnknownNotAMismatchWhenTheIdMatches(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-terra\"}");
+
+        assertEquals("ses_x", handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(),
+                "id matches and provider is simply unknown (absent), never a mismatch: " + exhausted);
+    }
+
+    /**
+     * The other half of the same fix: a missing {@code providerID} must NOT blind the check to a
+     * genuine id mismatch. This is what proves the fix narrows the comparison rather than switching
+     * the whole check off whenever {@code providerID} happens to be absent.
+     */
+    @Test
+    void aMissingProviderIdInTheEvidenceStillCatchesARealIdMismatch(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-sol\"}");
+
+        assertEquals("ses_x", handle.agentSessionId());
+        assertEquals(1, exhausted.size(),
+                "the id genuinely differs, so this must still quarantine even with providerID absent: "
+                        + exhausted);
+        assertTrue(exhausted.get(0).contains("gpt-5.6-sol") && exhausted.get(0).contains(cfg.model()),
+                "reports both the requested and actual model: " + exhausted.get(0));
+    }
+
+    /**
+     * THE TRAP, row 3: a profile that names no provider prefix (bare {@code "deepseek-v4-flash"})
+     * must match on id alone — the profile never asked for a specific provider, so opencode
+     * resolving it to {@code gx} is not evidence of anything wrong.
+     */
+    @Test
+    void aBareModelWithNoProviderPrefixMatchesOnIdAloneAndIsNotAMismatch(@TempDir Path configRoot,
+                                                                          @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("deepseek-v4-flash", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
+
+        assertEquals("ses_x", handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(),
+                "no provider was requested, so opencode's own provider resolution is not a mismatch: "
+                        + exhausted);
+    }
+
+    /**
+     * THE TRAP, row 4 (not a row in the table, the reason the table exists): a mismatched id, with
+     * the profile's requested and the actual model both named in the ERROR and the sink's reason —
+     * the exact fleetd #175 scenario (a withdrawn model silently falls back to a paid one).
+     */
+    @Test
+    void aRealIdMismatchLogsAnErrorNamingBothModelsAndQuarantinesThroughTheSink(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(target + "|" + reason);
+        FleetConfig.Profile cfg = opencodeCfg("opencode/nemotron-3-ultra-free", null, null);
+
+        Logger logger = (Logger) LoggerFactory.getLogger(OpenCodeLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        PeerHandle handle;
+        try {
+            handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                    .spawn(new SpawnRequest(null, "/work/dir", null));
+            OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                    "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+            assertEquals("ses_x", handle.agentSessionId());
+        } finally {
+            logger.detachAppender(appender);
+        }
+
+        assertEquals(1, exhausted.size(), "a real id mismatch must reach the sink exactly once");
+        assertTrue(exhausted.get(0).contains("opencode/nemotron-3-ultra-free"),
+                "the sink reason must name the requested model: " + exhausted.get(0));
+        assertTrue(exhausted.get(0).contains("gpt-5.6-sol"),
+                "the sink reason must name the actual model: " + exhausted.get(0));
+
+        List<ILoggingEvent> errors = appender.list.stream()
+                .filter(e -> e.getLevel() == Level.ERROR)
+                .toList();
+        assertEquals(1, errors.size(), "exactly one ERROR for the mismatch: " + appender.list);
+        String message = errors.get(0).getFormattedMessage();
+        assertTrue(message.contains("opencode/nemotron-3-ultra-free") && message.contains("gpt-5.6-sol")
+                        && message.contains(cfg.profile()),
+                "the ERROR must name the requested model, the actual model, AND the profile: " + message);
+
+        // The check runs at most once per handle even if agentSessionId() is polled again.
+        handle.agentSessionId();
+        assertEquals(1, exhausted.size(), "no duplicate quarantine on a repeated call");
+    }
+
+    /**
+     * fleetd #175's central safety rule: absent, empty, or unparseable model evidence is UNKNOWN,
+     * never a mismatch — it must never quarantine a working profile. Covers every "no real
+     * evidence yet" shape: no row at all, a row with a null model, and a row with unparseable JSON.
+     */
+    @Test
+    void unknownOrUnparseableModelEvidenceNeverQuarantines(@TempDir Path configRoot,
+                                                             @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+
+        // No row yet at all.
+        assertNull(handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(), "no row yet is UNKNOWN, not a mismatch: " + exhausted);
+
+        // A row exists (for a DIFFERENT directory) so a poll on ours still finds nothing.
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_other", "/work/other", 500L,
+                "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+        assertNull(handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(), "a row for a different directory is UNKNOWN, not a mismatch: "
+                + exhausted);
+    }
+
+    /**
+     * A profile with no {@code model:} configured has nothing to compare against — the check must
+     * stay silent no matter what opencode actually ran, since there is no requested value to be
+     * wrong about.
+     */
+    @Test
+    void aProfileWithNoConfiguredModelIsNeverCheckedForAMismatch(@TempDir Path configRoot,
+                                                                  @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg(null, null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"anything-at-all\",\"providerID\":\"anyone\"}");
+
+        assertEquals("ses_x", handle.agentSessionId());
+        assertTrue(exhausted.isEmpty(), "no model configured → nothing to compare: " + exhausted);
+    }
+
+    // --- fleetd #234 defect 1: key the model check on the RESOLVED id, not the shared directory --
+
+    /**
+     * The exact shape fleetd #234 reported: {@code fleet_spawn} with no {@code worktree:} shares
+     * the lead's cwd across every worker, so more than one session row can exist for the SAME
+     * {@code directory}. Once THIS handle's own session id is resolved, a sibling member spawned
+     * later into the same shared directory — writing a NEWER, unrelated row — must never make the
+     * already-resolved session look mismatched. Today's code re-derives "the newest row in this
+     * directory" on every call (both for the id AND, independently, for the model), so it would
+     * pick up the sibling's row on the second call and flag a false mismatch AND flip the returned
+     * id. The fix (fleetd #234) makes the id sticky once resolved and reads the model back for
+     * exactly that id (see {@link OpenCodeSessionDiscovery#actualModelForSessionId}) — never
+     * "whatever is newest in the directory right now."
+     */
+    @Test
+    void modelCheckReadsTheResolvedSessionsOwnRowNotWhateverIsNewestInTheSharedDirectory(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
+        PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
+                .spawn(new SpawnRequest(null, "/work/dir", null));
+
+        // Our own session's row, correctly matching the profile's requested model.
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_ours", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
+        assertEquals("ses_ours", handle.agentSessionId(), "resolves to our own session");
+        assertTrue(exhausted.isEmpty(), "matching model → no mismatch on first resolve: " + exhausted);
+
+        // A sibling member, spawned later into the SAME shared directory (no worktree, fleetd
+        // #234's default), writes a newer row running a totally different model.
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_sibling", "/work/dir", 9000L,
+                "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
+
+        assertEquals("ses_ours", handle.agentSessionId(),
+                "the session id, once resolved, must not flip to a sibling sharing the directory");
+        assertTrue(exhausted.isEmpty(),
+                "a sibling's later, unrelated row in the same shared directory must never be read "
+                        + "as OUR session's model: " + exhausted);
+    }
+
+    // --- fleetd #234 defect 2: the quarantine the ERROR announces must actually happen -----------
+
+    /**
+     * fleetd #234: {@link OpenCodeLauncher.SessionAwareHandle#checkModelMatch} fires from {@code
+     * agentSessionId()}, which {@code SessionManager.acquire()} calls to build the very first
+     * {@code MemberSession} record — BEFORE that session is put into the registry {@code
+     * sessions.roster()} reads. A sink that resolves {@code target -> profile} ONLY through the
+     * roster (today's {@code Fleetd.java} code, before this fix) therefore finds nothing at this
+     * exact moment and silently does not quarantine, even though it just logged an ERROR saying it
+     * would. This test drives the REAL path — {@code SessionManager.acquire()} — not the sink
+     * directly, because the bug is entirely about this ordering; a direct-sink test cannot see it
+     * (and is exactly why #175's own test suite, which only ever called the sink directly or after
+     * registration, never caught this).
+     *
+     * <p>The sink under test mirrors {@code Fleetd.main()}'s real wiring after the fix: resolve
+     * via the roster first (unchanged for {@code CompletionResolver}'s two call sites), falling
+     * back to the profile hint {@link OpenCodeLauncher} now supplies via {@link
+     * ExhaustionSink#onExhausted(String, String, String)} when the roster lookup misses.
+     */
+    @Test
+    void aSpawnTimeModelMismatchActuallyQuarantinesTheCredentialThroughTheRealAcquirePath(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        FleetConfig.Profile cfg = opencodeCfgWithCredential(
+                "terra", "opencode/nemotron-3-ultra-free", "openai-shared");
+        Map<String, FleetConfig.Profile> profiles = Map.of(cfg.profile(), cfg);
+        BackendQuarantine quarantine = new BackendQuarantine(() -> 0L, TimeUnit.SECONDS.toNanos(1800));
+
+        // fleetd #234, round 4: safely a lambda now — the 3-arg overload is the interface's single
+        // abstract method, so there is no 2-arg overload left to bind to instead.
+        ExhaustionSink sink = (target, reason, profileHint) -> {
+            FleetConfig.Profile profile = profileHint == null ? null : profiles.get(profileHint);
+            if (profile != null) {
+                quarantine.quarantine(profile.effectiveCredentialId());
+            }
+        };
+
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, sink);
+        SessionManager sessions = new SessionManager(launcher);
+
+        // The mismatching row exists BEFORE the spawn — reproducing fleetd #234's exact timing:
+        // opencode's session table already carries evidence by the moment acquire() first asks.
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+
+        assertFalse(quarantine.isQuarantined("openai-shared"), "nothing quarantined before the spawn");
+
+        // The real production entrypoint: acquire() builds the MemberSession by calling
+        // handle.agentSessionId() BEFORE registry.put() runs.
+        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+
+        assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
+        assertTrue(quarantine.isQuarantined("openai-shared"),
+                "the mismatch fires DURING acquire(), before roster registration, and must still "
+                        + "reach the quarantine via the profile hint — not silently no-op");
+    }
+
+    /**
+     * The other half of the same proof: a sink that resolves {@code target -> profile} ONLY
+     * through the roster (i.e. ignores the profile hint entirely, ~today's pre-fix {@code
+     * Fleetd.java}) drops the SAME spawn-time mismatch silently — the credential is never
+     * quarantined even though {@link OpenCodeLauncher} logged the mismatch ERROR. This is the
+     * failure fleetd #234 reported, reproduced through the real {@code SessionManager.acquire()}
+     * path rather than asserted by inspecting the fix.
+     */
+    @Test
+    void aRosterOnlySinkSilentlyDropsTheSpawnTimeQuarantine(@TempDir Path configRoot,
+                                                              @TempDir Path discRoot) throws Exception {
+        FleetConfig.Profile cfg = opencodeCfgWithCredential(
+                "terra", "opencode/nemotron-3-ultra-free", "openai-shared");
+        BackendQuarantine quarantine = new BackendQuarantine(() -> 0L, TimeUnit.SECONDS.toNanos(1800));
+
+        // Deliberately ignores the profile hint — the pre-fix shape: only a roster lookup (modelled
+        // here as always empty, since acquire() has not registered the session yet either way).
+        ExhaustionSink rosterOnlySink = (target, reason, profile) -> { };
+
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, rosterOnlySink);
+        SessionManager sessions = new SessionManager(launcher);
+
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+
+        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+
+        assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
+        assertFalse(quarantine.isQuarantined("openai-shared"),
+                "a roster-only sink cannot see this target yet — the quarantine silently never "
+                        + "happens, which is exactly fleetd #234 defect 2");
+    }
+
+    /**
+     * fleetd #234, round 2: the two tests above inject a sink DIRECTLY into {@link OpenCodeLauncher},
+     * which is not what {@code Fleetd.main} actually does. Production has one more hop: {@code
+     * Fleetd.java} builds the adapters (including {@code OpenCodeLauncher}) before {@code sessions}
+     * exists — a genuine construction-order cycle — so it hands the launcher a <em>forwarding</em>
+     * sink pointed at an {@code AtomicReference<ExhaustionSink>}, and only later {@code .set(...)}s
+     * that reference to the real sink once {@code sessions} is built.
+     *
+     * <p><strong>Round 3:</strong> a round-2 version of this test built its own copy of that
+     * forwarder's shape as an anonymous class. Mutating {@code Fleetd.java}'s ACTUAL forwarder back
+     * into a broken lambda left this test's own copy untouched, so it kept passing while production
+     * had regressed to the exact bug being fixed. This version instead calls {@link
+     * ExhaustionSink#forwardingTo}, the same factory {@code Fleetd.java} calls — the identical
+     * object, not a rebuilt copy of its shape — so a regression at either the {@code Fleetd.java}
+     * call site or inside {@code forwardingTo} itself has nowhere left to hide. See {@code
+     * ExhaustionSinkForwardingHazardTest} for the same factory exercised in isolation.
+     */
+    @Test
+    void theSpawnTimeQuarantineSurvivesTheFleetdStyleForwardingHop(@TempDir Path configRoot,
+                                                                     @TempDir Path discRoot) throws Exception {
+        FleetConfig.Profile cfg = opencodeCfgWithCredential(
+                "terra", "opencode/nemotron-3-ultra-free", "openai-shared");
+        Map<String, FleetConfig.Profile> profiles = Map.of(cfg.profile(), cfg);
+        BackendQuarantine quarantine = new BackendQuarantine(() -> 0L, TimeUnit.SECONDS.toNanos(1800));
+
+        // Fleetd.java:176 — the forwarding sink is built BEFORE the real one can exist, and the
+        // launcher below is constructed against this forwarder, exactly like Fleetd.main. Calling
+        // the SAME factory Fleetd.java calls — ExhaustionSink.forwardingTo — rather than rebuilding
+        // the forwarder's shape here is the whole point (fleetd #234, round 3): a round-2 version of
+        // this test built its own copy, so mutating Fleetd.java's real forwarder back into a lambda
+        // left this test untouched. Routed through the shared factory, a regression at either the
+        // Fleetd.java call site (reverting to a hand-written lambda) or inside the factory body
+        // itself has nowhere left to hide from this test.
+        java.util.concurrent.atomic.AtomicReference<ExhaustionSink> exhaustionSinkRef =
+                new java.util.concurrent.atomic.AtomicReference<>(ExhaustionSink.none());
+        ExhaustionSink forwardingExhaustionSink = ExhaustionSink.forwardingTo(exhaustionSinkRef::get);
+
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, forwardingExhaustionSink);
+        SessionManager sessions = new SessionManager(launcher);
+
+        // Fleetd.java:359-390 — the real sink is only built and pointed to AFTER `sessions` exists,
+        // same order as production. Safely a lambda (round 4): see the note on the forwarder above.
+        ExhaustionSink realSink = (target, reason, profileHint) -> {
+            FleetConfig.Profile profile = profileHint == null ? null : profiles.get(profileHint);
+            if (profile != null) {
+                quarantine.quarantine(profile.effectiveCredentialId());
+            }
+        };
+        exhaustionSinkRef.set(realSink);
+
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+
+        assertFalse(quarantine.isQuarantined("openai-shared"), "nothing quarantined before the spawn");
+
+        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+
+        assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
+        assertTrue(quarantine.isQuarantined("openai-shared"),
+                "the profile hint must survive the Fleetd-style forwarding hop and reach the real "
+                        + "sink — a lambda forwarder drops it and this must go red");
+    }
 }

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

@@ -27,19 +27,30 @@ class OpenCodeSessionDiscoveryTest {
      * and insert one row. Static so {@link OpenCodeLauncherTest} can reuse it.
      */
     static void writeRecord(Path root, String id, String directory, long timeUpdated) throws Exception {
+        writeRecord(root, id, directory, timeUpdated, null);
+    }
+
+    /**
+     * Same as {@link #writeRecord(Path, String, String, long)}, plus the {@code model} column
+     * fleetd #175 reads — the raw JSON opencode writes, e.g.
+     * {@code {"id":"gpt-5.6-terra","providerID":"openai"}}. {@code modelJson} may be {@code null}.
+     */
+    static void writeRecord(Path root, String id, String directory, long timeUpdated, String modelJson)
+            throws Exception {
         Path db = root.resolve("opencode.db");
         try (Connection connection = DriverManager.getConnection("jdbc:sqlite:" + db)) {
             try (Statement statement = connection.createStatement()) {
                 statement.execute("CREATE TABLE IF NOT EXISTS session ("
-                        + "id TEXT PRIMARY KEY, directory TEXT, time_updated INTEGER)");
+                        + "id TEXT PRIMARY KEY, directory TEXT, time_updated INTEGER, model TEXT)");
             }
             // Bound parameters, not string interpolation: the class under test uses a
             // PreparedStatement, and a hand-escaped INSERT here is a pattern someone copies out.
             try (PreparedStatement insert = connection.prepareStatement(
-                    "INSERT INTO session (id, directory, time_updated) VALUES (?, ?, ?)")) {
+                    "INSERT INTO session (id, directory, time_updated, model) VALUES (?, ?, ?, ?)")) {
                 insert.setString(1, id);
                 insert.setString(2, directory);
                 insert.setLong(3, timeUpdated);
+                insert.setString(4, modelJson);
                 insert.executeUpdate();
             }
         }
@@ -134,4 +145,107 @@ class OpenCodeSessionDiscoveryTest {
         assertNull(new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"),
                 "an unreadable database resolves to null, not an exception");
     }
+
+    // --- fleetd #175/#234: actualModelForSessionId / the model JSON column -----------------------
+
+    @Test
+    void parsesTheModelJsonIntoProviderAndId(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, "{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
+
+        OpenCodeSessionDiscovery.ActualModel actual =
+                new OpenCodeSessionDiscovery(root).actualModelForSessionId("ses_aaa");
+
+        assertNotNull(actual, "a well-formed model JSON parses");
+        assertEquals("gpt-5.6-terra", actual.id());
+        assertEquals("openai", actual.provider());
+    }
+
+    @Test
+    void looksUpByIdEvenWhenAnotherRowInTheSameDirectoryIsNewer(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_ours", "/work/dir", 1000L, "{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
+        writeRecord(root, "ses_sibling", "/work/dir", 9000L, "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
+
+        OpenCodeSessionDiscovery discovery = new OpenCodeSessionDiscovery(root);
+        assertEquals("gpt-5.6-terra", discovery.actualModelForSessionId("ses_ours").id(),
+                "querying by id reads OUR row, not the directory's newest row");
+        assertEquals("deepseek-v4-flash", discovery.actualModelForSessionId("ses_sibling").id(),
+                "each id resolves to its own row independently of time_updated ordering");
+    }
+
+    @Test
+    void anUnknownSessionIdYieldsUnknownModelRatherThanAMismatch(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, "{\"id\":\"x\",\"providerID\":\"y\"}");
+
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForSessionId("ses_no_such_row"),
+                "no row for this id yet → unknown, not a wrong model");
+    }
+
+    @Test
+    void aBlankOrNullSessionIdYieldsUnknownModel(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, "{\"id\":\"x\",\"providerID\":\"y\"}");
+
+        OpenCodeSessionDiscovery discovery = new OpenCodeSessionDiscovery(root);
+        assertNull(discovery.actualModelForSessionId(null));
+        assertNull(discovery.actualModelForSessionId("   "));
+    }
+
+    @Test
+    void aNullModelColumnYieldsUnknownWithoutThrowing(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, null);
+
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForSessionId("ses_aaa"),
+                "a row with no model value yet is unknown, not a mismatch");
+    }
+
+    @Test
+    void unparseableModelJsonYieldsUnknownWithoutThrowing(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, "this is not json");
+
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForSessionId("ses_aaa"),
+                "JSON that fails to parse resolves to unknown, never an exception");
+    }
+
+    @Test
+    void modelJsonMissingIdYieldsUnknown(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, "{\"providerID\":\"openai\"}");
+
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForSessionId("ses_aaa"),
+                "no id in the JSON → unknown, since id is what a caller actually compares");
+    }
+
+    @Test
+    void aMissingDatabaseYieldsUnknownModelWithoutThrowing(@TempDir Path root) {
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForSessionId("ses_aaa"));
+    }
+
+    /**
+     * fleetd #175 site 2: a schema that predates the {@code model} column entirely — the exact
+     * shape a real opencode upgrade/downgrade could produce. This must degrade to UNKNOWN for the
+     * model, and — the property that actually matters — must NOT take id resolution down with it.
+     * A combined single query for both columns would fail this test; that is why
+     * {@link OpenCodeSessionDiscovery#actualModelForSessionId} runs its own independent query.
+     */
+    @Test
+    void aMissingModelColumnYieldsUnknownButIdResolutionStillWorks(@TempDir Path root) throws Exception {
+        Path db = root.resolve("opencode.db");
+        try (Connection connection = DriverManager.getConnection("jdbc:sqlite:" + db)) {
+            try (Statement statement = connection.createStatement()) {
+                statement.execute("CREATE TABLE session (id TEXT PRIMARY KEY, directory TEXT, "
+                        + "time_updated INTEGER)");
+            }
+            try (PreparedStatement insert = connection.prepareStatement(
+                    "INSERT INTO session (id, directory, time_updated) VALUES (?, ?, ?)")) {
+                insert.setString(1, "ses_aaa");
+                insert.setString(2, "/w/a");
+                insert.setLong(3, 1000L);
+                insert.executeUpdate();
+            }
+        }
+
+        OpenCodeSessionDiscovery discovery = new OpenCodeSessionDiscovery(root);
+        assertEquals("ses_aaa", discovery.sessionIdForDirectory("/w/a"),
+                "id resolution must survive a database with no model column at all");
+        assertNull(discovery.actualModelForSessionId("ses_aaa"),
+                "no model column → unknown, not a throw and not a mismatch");
+    }
 }

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

@@ -1099,4 +1099,82 @@ class GitWorktreesTest {
         assertTrue(e.getMessage().contains("cb185-nonexistent-group-zz"),
                 "exception must name the missing/refused group: " + e.getMessage());
     }
+
+    // --- fleetd #224: worktreeRoot itself must be group-traversable, established in add() ---------
+
+    /**
+     * {@code add} must make the worktree ROOT itself group-traversable when a group is configured —
+     * established once here, where the root is created, never at a use site (never inside a
+     * launcher). A recording runner stands in for chgrp/chmod, the same seam
+     * {@link #shareWithGroupRunsConfigThenChgrpChmodSetgidPerPath} uses for the per-worktree share.
+     */
+    @Test
+    void addSharesWorktreeRootWithGroupWhenConfigured(@TempDir Path tmp) throws Exception {
+        Path repo = initRepo(tmp.resolve("repo"));
+        Path root = tmp.resolve("wts");
+        List<List<String>> recorded = new java.util.ArrayList<>();
+        java.util.function.Function<String[], String> recordingRunner = cmd -> {
+            recorded.add(joined(cmd));
+            return "";
+        };
+        GitWorktrees gitWorktrees = new GitWorktrees(root.toString(), "devteam", _ -> {}, recordingRunner);
+
+        gitWorktrees.add(repo.toString(), "cb-224-branch", "HEAD");
+
+        assertTrue(recorded.contains(List.of("chgrp", "devteam", root.toString())),
+                "the worktree root itself must be chgrp'd to the configured group: " + recorded);
+        assertTrue(recorded.contains(List.of("chmod", "g+x", root.toString())),
+                "the worktree root itself must gain group-execute so a different-uid member can "
+                        + "traverse into it: " + recorded);
+    }
+
+    /** {@code worktreeGroup} unset (today's only live mode) ⇒ {@code add} spawns no share process
+     *  for the root at all — behaviour must be byte-identical to before fleetd #224. */
+    @Test
+    void addSharesNothingForTheRootWhenNoGroupConfigured(@TempDir Path tmp) throws Exception {
+        Path repo = initRepo(tmp.resolve("repo"));
+        Path root = tmp.resolve("wts");
+        List<List<String>> recorded = new java.util.ArrayList<>();
+        java.util.function.Function<String[], String> recordingRunner = cmd -> {
+            recorded.add(joined(cmd));
+            return "";
+        };
+        GitWorktrees gitWorktrees = new GitWorktrees(root.toString(), null, _ -> {}, recordingRunner);
+
+        gitWorktrees.add(repo.toString(), "cb-224-nogroup", "HEAD");
+
+        assertTrue(recorded.isEmpty(), "no group configured must spawn no share process for the "
+                + "root at all: " + recorded);
+    }
+
+    /**
+     * fleetd #224, acceptance criterion 2/3: when the root cannot be made group-traversable — here
+     * because the configured group does not exist, the same real-failure shape
+     * {@link #shareWithGroupThrowsNamingTheGroupWhenChgrpFails} drives for the per-worktree share —
+     * the spawn is refused with a message naming the root, its current mode, and the group. This
+     * drives the REAL {@code chgrp} (no recording runner), and the refusal happens before {@code git
+     * worktree add} ever runs, so no partial worktree is left behind either.
+     */
+    @Test
+    void addRefusesWhenWorktreeRootCannotBeMadeGroupTraversable(@TempDir Path tmp) throws Exception {
+        Path repo = initRepo(tmp.resolve("repo"));
+        Path root = tmp.resolve("wts");
+        GitWorktrees gitWorktrees = new GitWorktrees(root.toString(), "cb224-nonexistent-group-zz");
+
+        WorktreeException e = assertThrows(WorktreeException.class,
+                () -> gitWorktrees.add(repo.toString(), "cb-224-refuse", "HEAD"));
+
+        assertTrue(e.getMessage().contains(root.toString()),
+                "refusal must name the worktree root: " + e.getMessage());
+        assertTrue(e.getMessage().contains("cb224-nonexistent-group-zz"),
+                "refusal must name the missing/refused group: " + e.getMessage());
+        assertTrue(Files.isDirectory(root), "the root is created before the group check runs");
+        String mode = java.nio.file.attribute.PosixFilePermissions.toString(Files.getPosixFilePermissions(root));
+        assertTrue(e.getMessage().contains(mode),
+                "refusal must name the root's current mode (" + mode + "): " + e.getMessage());
+        try (java.util.stream.Stream<Path> children = Files.list(root)) {
+            assertTrue(children.findAny().isEmpty(), "no worktree must be left behind under the root: "
+                    + "the refusal must happen before `git worktree add` ever runs");
+        }
+    }
 }

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

@@ -4,6 +4,8 @@ import ch.qos.logback.classic.Level;
 import ch.qos.logback.classic.LoggerContext;
 import ch.qos.logback.classic.spi.ILoggingEvent;
 import ch.qos.logback.core.read.ListAppender;
+import dev.ltms.fleet.auth.MemberRegistry;
+import dev.ltms.fleet.auth.MemberLifecycle;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.guard.SubscriptionGuard;
 import dev.ltms.fleet.herdr.AgentControl;
@@ -333,6 +335,110 @@ class SessionManagerTest {
         }
     }
 
+    /**
+     * fleetd #226: a contended slot is refused through the real {@link SessionManager#acquire}
+     * path before the real launcher can hand an architect charter to a process.
+     */
+    @Test
+    void aContendedArchitectSlotRefusesBeforeTheLauncherStartsAMember() {
+        FakeHerdr herdr = new FakeHerdr();
+        SessionManager sessions = sessionManager(herdr);
+        MemberRegistry members = architectRegistry();
+        assertTrue(members.bind("architect:opus", "term_already_bound"),
+                "precondition: occupy the sole architect slot before the real spawn under test");
+        sessions.setMemberLifecycle(members);
+
+        assertThrows(IllegalArgumentException.class, () -> sessions.acquire("ltms-local", MemberRole.ARCHITECT,
+                null, "/caller", "term_primary", null));
+
+        assertFalse(herdr.called("agent.start"), "a refused slot must never reach the launcher");
+        assertTrue(sessions.roster().isEmpty(), "no member exists to receive the wrong charter");
+    }
+
+    @Test
+    void aSecondArchitectOnAnAlreadyBoundProfileIsHeldAsDevNotArchitectAndWarnsLoudly() {
+        FakeHerdr herdr = new FakeHerdr();
+        SessionManager sessions = sessionManager(herdr);
+        MemberRegistry members = architectRegistry();
+        sessions.setMemberLifecycle(bindFailureAfterReservation(members));
+        LoggerContext ctx = (LoggerContext) LoggerFactory.getILoggerFactory();
+        ch.qos.logback.classic.Logger registryLog = (ch.qos.logback.classic.Logger)
+                LoggerFactory.getLogger(MemberRegistry.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.setContext(ctx);
+        appender.start();
+        registryLog.addAppender(appender);
+        registryLog.setLevel(Level.WARN);
+        try {
+            MemberSession session = sessions.acquire("ltms-local", MemberRole.ARCHITECT, null,
+                    "/caller", "term_primary", null);
+
+            assertEquals(MemberRole.DEV, session.role(), "a failed reservation bind must use the fallback");
+            String warn = appender.list.stream()
+                    .filter(e -> e.getLevel().equals(Level.WARN))
+                    .map(ILoggingEvent::getFormattedMessage)
+                    .findFirst()
+                    .orElse("no slot-exhaustion WARN logged");
+            assertTrue(warn.contains("ltms-local"), "the WARN names the profile: " + warn);
+            assertTrue(warn.contains(session.terminalId()), "the WARN names the terminal: " + warn);
+        } finally {
+            registryLog.detachAppender(appender);
+        }
+    }
+
+    @Test
+    void reservedArchitectSlotBindsTheLaunchedMember() {
+        FakeHerdr herdr = new FakeHerdr();
+        SessionManager sessions = sessionManager(herdr);
+        sessions.setMemberLifecycle(architectRegistry());
+
+        MemberSession session = sessions.acquire("ltms-local", MemberRole.ARCHITECT, null,
+                "/caller", "term_primary", null);
+
+        assertEquals(MemberRole.ARCHITECT, session.role());
+    }
+
+    private static MemberRegistry architectRegistry() {
+        return new MemberRegistry(new FleetConfig.Fleet(Map.of(), Map.of("opus", new FleetConfig.Slot("ltms-local")),
+                Map.of(), Map.of(), null));
+    }
+
+    private static MemberLifecycle bindFailureAfterReservation(MemberRegistry members) {
+        return new MemberLifecycle() {
+            @Override
+            public MemberRole acquired(MemberRole role, String profile, String terminal) {
+                return members.acquired(role, profile, terminal);
+            }
+
+            @Override
+            public void released(String terminal) {
+                members.released(terminal);
+            }
+
+            @Override
+            public void requireSlotFor(MemberRole role, String profile) {
+                members.requireSlotFor(role, profile);
+            }
+
+            @Override
+            public SlotReservation reserve(MemberRole role, String profile) {
+                return members.reserve(role, profile);
+            }
+
+            @Override
+            public boolean bind(SlotReservation reservation, String terminal) {
+                members.release(reservation);
+                assertTrue(members.bind(reservation.slot(), "term_racer"), "the racer takes the released slot");
+                return false;
+            }
+
+            @Override
+            public void release(SlotReservation reservation) {
+                members.release(reservation);
+            }
+        };
+    }
+
     @Test
     void rosterReflectsAcquiredMinusReleased() {
         FakeHerdr herdr = new FakeHerdr();
@@ -599,6 +705,32 @@ class SessionManagerTest {
                 "no session is registered when spawn times out (roster empty)");
     }
 
+    @Test
+    void failedArchitectLaunchReleasesItsReservationForTheNextLaunch() {
+        FakeHerdr herdr = new FakeHerdr();
+        herdr.agentStatus("unknown");
+        long[] clock = {0};
+        FleetConfig.Profile cfg = new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
+                List.of("ccs", "ltms-local"), "tab", "fleetd-workers",
+                "worker: {profile} #{n}", null, null, null);
+        ClaudeCodeLauncher workers = new ClaudeCodeLauncher(
+                new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null,
+                1, () -> clock[0], () -> clock[0] += 10);
+        SessionManager sessions = new SessionManager(workers, new GitWorktrees(), () -> 0L, 0);
+        sessions.setMemberLifecycle(architectRegistry());
+
+        assertThrows(PeerUnreachableException.class, () -> sessions.acquire("ltms-local", MemberRole.ARCHITECT,
+                null, "/caller", "term_primary", null));
+
+        herdr.agentStatus("idle");
+        MemberSession retry = sessions.acquire("ltms-local", MemberRole.ARCHITECT, null,
+                "/caller", "term_primary", null);
+        assertEquals(MemberRole.ARCHITECT, retry.role(),
+                "the failed launch returned its reservation instead of silently shrinking the slot pool");
+    }
+
     // --- CB-516: release must notify, so a blocked send can be failed --------------------------
 
     @Test

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

@@ -95,10 +95,9 @@ class WorktreeSessionManagerTest {
                 null, "/caller/proj", null, null);
         assertEquals("architect:architect", members.slotForTerminal(replacement.terminalId()));
 
-        MemberSession overflow = sessions.acquire("ltms-local", MemberRole.ARCHITECT,
-                null, "/caller/proj", null, null);
-        assertNull(members.slotForTerminal(overflow.terminalId()),
-                "a full slot pool must not stop the architect spawn");
+        assertThrows(IllegalArgumentException.class, () -> sessions.acquire("ltms-local", MemberRole.ARCHITECT,
+                null, "/caller/proj", null, null),
+                "a full slot pool must stop the spawn before it can receive an architect charter");
     }
 
     @Test