fleetd#201 Unit 1: typed backend-error classification in CompletionResolver

Replace the hardcoded API-Error check with a target-keyed BackendErrorPatternLookup
plus a BackendErrorSink, mirroring the existing ExhaustedPatternLookup/ExhaustionSink
pair. Classifies in all three paths (normal block, #211 raw-scrape fallback, and the
fleetd#164 MIN_TURN_NANOS floor). The sink fires only after Rendezvous.resolveFailure
wins for the exact waiter. A target with no configured pattern still falls back to the
narrow (?i)\bAPI Error\s*: compatibility pattern. Existing constructors keep compiling
via BackendErrorPatternLookup.legacy() / BackendErrorSink.none() defaults.

Public send result is unchanged (still a failed send) — the typed sink event is the
internal seam Unit 5 will consume.

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

@@ -169,6 +169,12 @@ 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());
+        ExhaustionSink forwardingExhaustionSink = (target, reason) -> exhaustionSinkRef.get().onExhausted(target, reason);
         // 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 +190,7 @@ 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
@@ -334,6 +340,11 @@ 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.
+        //
+        // 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.
         ExhaustionSink exhaustionSink = (target, reason) -> sessions.roster().stream()
                 .filter(session -> target.equals(session.terminalId()))
                 .findFirst()
@@ -342,10 +353,12 @@ public final class Fleetd {
                 .ifPresent(profile -> {
                     String credentialId = profile.effectiveCredentialId();
                     quarantine.quarantine(credentialId);
-                    log.warn("credential '{}' quarantined for {}s (profile '{}' classified "
-                                    + "BACKEND_EXHAUSTED): {}", 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/BackendErrorPatternLookup.java

@@ -0,0 +1,35 @@
+package dev.ltms.fleet.inject;
+
+import java.util.regex.Pattern;
+
+/**
+ * Per-target lookup for a profile's configured backend-error pattern (fleetd#201 / #227): how
+ * {@link CompletionResolver} recognizes a backend that failed a turn outright (a credential
+ * outage, a provider 5xx) from whatever ends up on the member's pane, apart from a genuine
+ * completion.
+ *
+ * <p>The pattern is always profile config, never a vendor string in Java source: every backend
+ * words its failure differently, so a hardcoded sentence would only ever match one of them. A
+ * target with no configured pattern is not "off" the way {@link ExhaustedPatternLookup#none()}
+ * is — {@link CompletionResolver} falls back to its narrow {@code (?i)\bAPI Error\s*:}
+ * compatibility pattern instead, so classification still happens, just without a profile-specific
+ * match. {@link #legacy()} is the explicit stand-in existing (pre-fleetd#201) callers pass to get
+ * exactly that fallback-only behavior until a caller wires a real, profile-driven lookup
+ * (fleetd#201 Unit 5).
+ */
+@FunctionalInterface
+public interface BackendErrorPatternLookup {
+
+    /** The compiled pattern configured for {@code target}'s profile, or {@code null} if none. */
+    Pattern patternFor(String target);
+
+    /**
+     * Legacy lookup — no profile has a configured pattern, so {@link CompletionResolver} classifies
+     * every target using only its built-in {@code (?i)\bAPI Error\s*:} compatibility pattern. The
+     * explicit stand-in existing constructors pass so their behavior is unchanged until a caller
+     * wires a real, profile-driven lookup.
+     */
+    static BackendErrorPatternLookup legacy() {
+        return target -> null;
+    }
+}

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

@@ -0,0 +1,35 @@
+package dev.ltms.fleet.inject;
+
+/**
+ * Notified when {@link CompletionResolver} actually delivers a typed backend-error classification
+ * to a waiting send (fleetd#201 / #227) — never on a race that lost. {@link CompletionResolver}
+ * calls this only after {@code Rendezvous.resolveFailure} returns {@code true} for that exact
+ * waiter, mirroring the win-only race rule {@link ExhaustionSink} already uses.
+ *
+ * <p>The public send result is unchanged by this classification — it is still a failed send
+ * ({@code Rendezvous.Kind#FAILED}); this sink is the internal seam a later stage (fleetd#201 Unit
+ * 5, the policy that cools a credential off after two such failures in 60 seconds and tells the
+ * lead) consumes. {@link CompletionResolver} knows only {@code target} (a herdr terminal id); it
+ * has no notion of profiles or credentials, so mapping {@code target} to whatever should be
+ * quarantined is entirely the sink's job — exactly like {@link ExhaustionSink}.
+ */
+@FunctionalInterface
+public interface BackendErrorSink {
+
+    /**
+     * @param target      the herdr terminal id whose turn was classified as a backend error
+     * @param matchedLine the single pane line that matched the backend-error pattern
+     * @param reason      the full failure reason carried by the classification (the matched line
+     *                    plus whatever pane context the classifying path attaches)
+     */
+    void onBackendError(String target, String matchedLine, String reason);
+
+    /**
+     * Inert sink — nothing happens on a backend-error classification. The explicit stand-in a
+     * caller (or a test not exercising this feature) passes instead of a defaulting overload,
+     * exactly like {@link ExhaustionSink#none()}.
+     */
+    static BackendErrorSink none() {
+        return (target, matchedLine, reason) -> { };
+    }
+}

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

@@ -89,6 +89,8 @@ public final class CompletionResolver implements TurnListener {
     private final Rendezvous rendezvous;
     private final ExhaustedPatternLookup exhaustedPatterns;
     private final ExhaustionSink exhaustionSink;
+    private final BackendErrorPatternLookup backendErrorPatterns;
+    private final BackendErrorSink backendErrorSink;
     private final LongSupplier nowNanos;
 
     /**
@@ -129,10 +131,48 @@ public final class CompletionResolver implements TurnListener {
      *                          classification actually resolves a waiter. Required for the same
      *                          reason as {@code exhaustedPatterns} — pass {@link ExhaustionSink#none()}
      *                          to opt out.
+     *
+     * <p>Transition constructor (fleetd#201 Unit 1): delegates to the full constructor below with
+     * {@link BackendErrorPatternLookup#legacy()} and {@link BackendErrorSink#none()}, so every
+     * existing caller keeps today's behavior — the narrow built-in {@code API Error:} match, no
+     * sink notified — until a caller wires a real, profile-driven backend-error lookup and sink
+     * (fleetd#201 Unit 5).
      */
     public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
                               ExhaustionSink exhaustionSink) {
-        this(agents, rendezvous, exhaustedPatterns, exhaustionSink, System::nanoTime);
+        this(agents, rendezvous, exhaustedPatterns, exhaustionSink,
+                BackendErrorPatternLookup.legacy(), BackendErrorSink.none(), System::nanoTime);
+    }
+
+    /**
+     * Transition constructor (fleetd#201 Unit 1): same legacy backend-error defaults as the 4-arg
+     * constructor above, but with the injectable clock. Kept so existing fleetd#164 timing tests
+     * (and any other caller that wants a controllable clock but not the backend-error feature)
+     * keep compiling unchanged.
+     */
+    public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
+                       ExhaustionSink exhaustionSink, LongSupplier nowNanos) {
+        this(agents, rendezvous, exhaustedPatterns, exhaustionSink,
+                BackendErrorPatternLookup.legacy(), BackendErrorSink.none(), nowNanos);
+    }
+
+    /**
+     * Production constructor (fleetd#201 Unit 1): adds the target-keyed backend-error pattern
+     * lookup and sink alongside the existing exhaustion pair. Required, like {@code exhaustedPatterns}
+     * and {@code exhaustionSink} — pass {@link BackendErrorPatternLookup#legacy()} and
+     * {@link BackendErrorSink#none()} to opt out.
+     *
+     * @param backendErrorPatterns fleetd#201: per-target lookup for a profile's configured
+     *                             backend-error pattern; a target with none configured is matched
+     *                             against the built-in compatibility pattern instead (never "off").
+     * @param backendErrorSink     fleetd#201: notified when a backend-error classification actually
+     *                             resolves a waiter — never on a race that lost.
+     */
+    public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
+                       ExhaustionSink exhaustionSink, BackendErrorPatternLookup backendErrorPatterns,
+                       BackendErrorSink backendErrorSink) {
+        this(agents, rendezvous, exhaustedPatterns, exhaustionSink, backendErrorPatterns, backendErrorSink,
+                System::nanoTime);
     }
 
     /**
@@ -145,11 +185,14 @@ public final class CompletionResolver implements TurnListener {
      * package.
      */
     public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
-                       ExhaustionSink exhaustionSink, LongSupplier nowNanos) {
+                       ExhaustionSink exhaustionSink, BackendErrorPatternLookup backendErrorPatterns,
+                       BackendErrorSink backendErrorSink, LongSupplier nowNanos) {
         this.agents = agents;
         this.rendezvous = rendezvous;
         this.exhaustedPatterns = Objects.requireNonNull(exhaustedPatterns, "exhaustedPatterns");
         this.exhaustionSink = Objects.requireNonNull(exhaustionSink, "exhaustionSink");
+        this.backendErrorPatterns = Objects.requireNonNull(backendErrorPatterns, "backendErrorPatterns");
+        this.backendErrorSink = Objects.requireNonNull(backendErrorSink, "backendErrorSink");
         this.nowNanos = Objects.requireNonNull(nowNanos, "nowNanos");
     }
 
@@ -232,16 +275,18 @@ public final class CompletionResolver implements TurnListener {
         // screen, since that is usually the backend's own error.
         long elapsedNanos = nowNanos.getAsLong() - turn.deliveredAtNanos();
         if (elapsedNanos < MIN_TURN_NANOS) {
-            fail(target, turn, tooFastReason(target, elapsedNanos));
+            failTooFast(target, turn, waiter, elapsedNanos);
             return;
         }
         String tail;
         String assistantBlock = null;
+        String rawScrape = null;
         int originalLength = 0;
         boolean clipped = false;
         boolean scrapeFailed = false;
         try {
-            assistantBlock = lastAssistantBlock(agents.read(target, SCRAPE_SOURCE));
+            rawScrape = agents.read(target, SCRAPE_SOURCE);
+            assistantBlock = lastAssistantBlock(rawScrape);
             originalLength = assistantBlock.strip().length();
             clipped = originalLength > MAX_SCRAPE_CHARS;
             tail = clip(assistantBlock);
@@ -256,6 +301,20 @@ public final class CompletionResolver implements TurnListener {
         // member, so a caller (including a lead deciding whether to delegate again) can tell a lost
         // turn from a real empty answer.
         if (scrapeFailed || tail.isEmpty()) {
+            // fleetd#211: lastAssistantBlock() found nothing usable — most often a pane with no ⏺
+            // marker at all, whose boundary scan then starts at the top of the raw screen and breaks
+            // immediately on the first line of TUI chrome (╭, │, ❯, …). Before giving up as a lost
+            // turn, run the same exhaustion/backend-error classification against the RAW scrape as a
+            // fallback, ONLY here. A pane that already yielded a usable block never reaches this
+            // branch, so the narrow (trimmed) match on the normal path below is completely unchanged
+            // — zero new false positives there. Every pane this fallback examines was already headed
+            // for the empty-scrape failure, so a wrong label here is strictly less bad than silently
+            // losing an exhaustion signal: the alternative outcome is already a failure, just one that
+            // never quarantines the credential. lastAssistantBlock stays the source of the reply
+            // TEXT everywhere else; only classification ever consults the raw scrape, and only here.
+            if (rawScrape != null && classifyRawScrapeFallback(target, turn, waiter, rawScrape)) {
+                return;
+            }
             fail(target, turn, emptyScrapeReason(target, scrapeFailed));
             return;
         }
@@ -286,18 +345,27 @@ public final class CompletionResolver implements TurnListener {
             }
             return;
         }
-        // fleetd#164 (part 2): a scrape that read cleanly and produced content still isn't a real
-        // reply when that content is the backend's own rejection (e.g. an HTTP 400 before the worker
-        // did any work). Classify it as a failure naming the member, rather than handing the caller a
-        // scrape that reads like a completed answer.
-        String backendError = firstMatchingLine(assistantBlock, BACKEND_ERROR);
+        // fleetd#164 (part 2) / fleetd#201: a scrape that read cleanly and produced content still
+        // isn't a real reply when that content is the backend's own rejection (e.g. an HTTP 400
+        // before the worker did any work). Classify it as a failure naming the member, rather than
+        // handing the caller a scrape that reads like a completed answer, and — only on the
+        // resolution that actually wins the race, mirroring the exhaustion sink above — notify the
+        // typed backend-error sink so a later stage can act on repeated failures.
+        String backendError = firstMatchingLine(assistantBlock, backendErrorPatternOrFallback(target));
         if (backendError != null) {
             // Carry the whole scrape, not just the matched line. The pattern is a heuristic: a member
             // that forgot fleet_reply while reporting *about* a backend error matches it too. Failing
             // is still right — the caller must not read a scrape as an answer — but dropping the rest
             // of the pane would destroy the report, which is the same defect fleetd#164 is about.
-            fail(target, turn, "member " + target + " ended on a backend error: " + backendError
-                    + "\n--- pane tail ---\n" + tail);
+            String reason = "member " + target + " ended on a backend error: " + backendError
+                    + "\n--- pane tail ---\n" + tail;
+            if (rendezvous.resolveFailure(waiter, reason)) {
+                inFlight.remove(target, turn);
+                log.warn("failing send to {} via turn-stall fallback: {}", target, reason);
+                // fleetd#201 Unit 1: only on the resolution that actually won the race — a late
+                // duplicate must never double-count one backend failure.
+                backendErrorSink.onBackendError(target, backendError, reason);
+            }
             return;
         }
         String completion = clipped ? tail + "\n" + CLIPPED_PANE_TAIL_MARKER : tail;
@@ -313,6 +381,49 @@ public final class CompletionResolver implements TurnListener {
         }
     }
 
+    /**
+     * fleetd#211: the raw-scrape fallback classification, run only when {@link #lastAssistantBlock}
+     * found nothing usable (see the call site in {@link #resolve}). Mirrors the two classifications
+     * the normal path already applies to the trimmed assistant block — exhaustion first, then the
+     * narrow {@link #BACKEND_ERROR} pattern — against {@code raw} instead, and reports whether one of
+     * them handled the turn (resolved the waiter or failed it) so the caller skips the empty-scrape
+     * failure. Never runs on the normal (non-empty-block) path, and never touches the reply text.
+     */
+    private boolean classifyRawScrapeFallback(String target, InFlight turn,
+                                               CompletableFuture<Rendezvous.Resolution> waiter, String raw) {
+        Pattern exhausted = exhaustedPatterns.patternFor(target);
+        String matchedLine = exhausted == null ? null : firstMatchingLine(raw, exhausted);
+        if (matchedLine != null) {
+            String reason = "backend exhausted (usage limit): " + matchedLine;
+            if (rendezvous.resolveExhausted(waiter, reason)) {
+                inFlight.remove(target, turn);
+                log.warn("completion for {} classified BACKEND_EXHAUSTED from the raw scrape (no "
+                        + "usable assistant block; no fleet_reply): {}", target, reason);
+                // CB-578 stage B: only on the resolution that actually won the race — a late
+                // duplicate must never quarantine a credential twice for one refusal.
+                exhaustionSink.onExhausted(target, reason);
+            }
+            return true;
+        }
+        String backendError = firstMatchingLine(raw, backendErrorPatternOrFallback(target));
+        if (backendError != null) {
+            // Carry the pane, not just the matched line — the same fleetd#164 rule the normal path
+            // above applies. Here it matters more, not less: the trimmed block was empty, so the raw
+            // scrape is the ONLY copy of whatever the member managed to say. Clipped to the same cap
+            // the normal path uses, since a raw screen has no boundary trimming to bound it.
+            String reason = "member " + target + " ended on a backend error: " + backendError
+                    + "\n--- pane tail ---\n" + clip(raw);
+            if (rendezvous.resolveFailure(waiter, reason)) {
+                inFlight.remove(target, turn);
+                log.warn("failing send to {} via turn-stall fallback from the raw scrape: {}", target, reason);
+                // fleetd#201 Unit 1: only on the resolution that actually won the race.
+                backendErrorSink.onBackendError(target, backendError, reason);
+            }
+            return true;
+        }
+        return false;
+    }
+
     /** Synchronous fail (the unit-testable core of {@link #onTurnFailed}). */
     void fail(String target, InFlight turn) {
         fail(target, turn, null);
@@ -349,22 +460,53 @@ public final class CompletionResolver implements TurnListener {
     }
 
     /**
-     * fleetd#164: the failure reason for a turn that settled inside {@link #MIN_TURN_NANOS} — names
-     * the member and both timings, and appends whatever the pane shows (usually the backend's own
-     * error) so the caller sees the cause, not just "it failed".
+     * fleetd#164 (floor) / fleetd#201 (classification): fail a turn that settled inside
+     * {@link #MIN_TURN_NANOS} — a crash signature (e.g. a backend HTTP 400 before the worker did
+     * anything) that a bare {@code BUSY -> DONE} transition cannot be told apart from a genuinely
+     * fast completion. Runs the same backend-error classification the normal and raw-scrape paths
+     * apply, against whatever is on screen right now: a match is a typed failure that notifies
+     * {@link #backendErrorSink} (only on the resolution that wins the race); a non-match stays the
+     * original generic too-fast failure, naming the member and both timings, with whatever the pane
+     * shows appended so the caller sees the cause, not just "it failed".
      */
-    private String tooFastReason(String target, long elapsedNanos) {
+    private void failTooFast(String target, InFlight turn, CompletableFuture<Rendezvous.Resolution> waiter,
+                              long elapsedNanos) {
         String scrape;
         try {
-            scrape = clip(agents.read(target, SCRAPE_SOURCE));
+            scrape = agents.read(target, SCRAPE_SOURCE);
         } catch (RuntimeException e) {
             scrape = "";
         }
-        String reason = String.format(
+        String clippedScrape = clip(scrape);
+        String baseReason = String.format(
                 "member %s went BUSY -> DONE in %dms (floor %dms) — too fast to be real work, most "
                         + "likely a backend error before any work started",
                 target, elapsedNanos / 1_000_000, MIN_TURN_NANOS / 1_000_000);
-        return scrape.isBlank() ? reason : reason + ": " + scrape;
+        String backendError = firstMatchingLine(scrape, backendErrorPatternOrFallback(target));
+        if (backendError != null) {
+            String reason = baseReason + ": " + clippedScrape;
+            if (rendezvous.resolveFailure(waiter, reason)) {
+                inFlight.remove(target, turn);
+                log.warn("failing send to {} via turn-stall fallback: {}", target, reason);
+                // fleetd#201 Unit 1: only on the resolution that actually won the race.
+                backendErrorSink.onBackendError(target, backendError, reason);
+            }
+            return;
+        }
+        fail(target, turn, clippedScrape.isBlank() ? baseReason : baseReason + ": " + clippedScrape);
+    }
+
+    /**
+     * fleetd#201: the pattern to classify a backend error against for {@code target} — its
+     * profile's configured {@link BackendErrorPatternLookup} entry when there is one, else the
+     * built-in {@link #BACKEND_ERROR} compatibility pattern. A target with no configured pattern is
+     * never "off": it always falls back to this narrow default, exactly like the classification
+     * behaved before fleetd#201 (Unit 5 reports a target relying on this fallback separately, as
+     * legacy-default coverage rather than full per-profile coverage).
+     */
+    private Pattern backendErrorPatternOrFallback(String target) {
+        Pattern configured = backendErrorPatterns.patternFor(target);
+        return configured != null ? configured : BACKEND_ERROR;
     }
 
     /**

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

@@ -278,18 +278,22 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
 
     /**
      * Add the Claude-specific session-identity flags to {@code argv} and return the peer's OWN
-     * session id — the resume handle. A resume request passes the prior id via {@code -r} and
-     * returns that id; a fresh named session mints a new UUID, passes it via {@code --session-id},
-     * and returns the mint. The bridge's logical name rides along as {@code -n} when present. When
-     * <em>no</em> identity is requested (sessionName and resumeSessionId both blank) this adds
-     * nothing and returns {@code null}, keeping the legacy no-identity launch byte-identical.
+     * session id — the resume handle. A resume passes the prior id via {@code -r} and returns
+     * that id; every other spawn mints a new UUID, passes it via {@code --session-id}, and
+     * returns the mint. The bridge's logical name rides along as {@code -n} when present.
+     *
+     * <p>fleetd #214: the mint is unconditional. A plain {@code fleet_spawn} passes neither
+     * sessionName nor resumeSessionId, yet the member must still be resumable, and this id is
+     * the only resume handle a claude-code member has — unlike opencode, nothing resolves it
+     * after the launch. Checked against the real binary (claude 2.1.252): the flag is safe on
+     * every spawn. The binary takes only a valid UUID — it refuses any other value at argument
+     * parsing ("Invalid session ID. Must be a valid UUID.") — so the {@code UUID.randomUUID()}
+     * mint is required, not incidental. The only flag interaction the binary documents is with
+     * {@code -r} (both claim the session id), and the resume branch above never combines the two.
      */
     private static String applySessionIdentity(List<String> argv, String sessionName, String resumeSessionId) {
         boolean resuming = resumeSessionId != null && !resumeSessionId.isBlank();
         boolean named = sessionName != null && !sessionName.isBlank();
-        if (!resuming && !named) {
-            return null; // no identity requested — keep the legacy launch byte-identical
-        }
         if (named) {
             argv.add("-n");
             argv.add(sessionName);
@@ -320,11 +324,19 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
      *
      * <p>CB-618: Claude Code refuses to start when BOTH {@code --append-system-prompt} and
      * {@code --append-system-prompt-file} are on the command line ("Cannot use both ... Please use
-     * only one"), so the two charters can never travel on separate flags. When both are present they
-     * are concatenated into the one file, role charter first and reply charter last — last is where
-     * the reply rule must sit, because it is the rule that must survive. When only the reply charter
-     * is present it keeps its proven inline {@code --append-system-prompt} delivery, which is also
-     * the only form that reaches a member with no repo checkout.
+     * only one"), so the two charters can never travel on separate flags. They are concatenated
+     * into the one file, role charter first and reply charter last — last is where the reply rule
+     * must sit, because it is the rule that must survive.
+     *
+     * <p>fleetd #220: a lone reply charter used to ride inline on {@code --append-system-prompt},
+     * which put ~800 bytes of prose on the command line herdr types into the pane. That line is
+     * capped at {@value HerdrPeerLauncher#PANE_COMMAND_BYTE_LIMIT} bytes by the pty itself, and
+     * everything past the cap is dropped with no error from any layer. The charter alone left about
+     * 50 bytes of headroom, so adding one flag ({@code --session-id}, fleetd #214) truncated the
+     * LAST argument instead — {@code --autocompact 250000} arrived as {@code --autocompact 25},
+     * claude rejected it, and every claude-code spawn died as an unexplained readiness timeout.
+     * The charter now always travels as a file, which takes the prose off the command line for
+     * good; {@link HerdrPeerLauncher#checkPaneCommandFits} is the backstop for whatever grows next.
      */
     private List<String> argvWithFleet(FleetConfig.Profile cfg, LaunchSpec spec) {
         String roleCharter = nonBlank(spec.roleCharter());
@@ -341,16 +353,13 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
             argv.add("--mcp-config");
             argv.add(mcpConfigJson(cfg));
         }
-        // Combine the charters in order role -> reply, dropping any that are absent. When two
-        // or more survive they must ride one --append-system-prompt-file (CB-618 forbids the inline
-        // flag and the file flag together). A lone reply charter keeps its proven inline delivery.
+        // Combine the charters in order role -> reply, dropping any that are absent. They ride one
+        // --append-system-prompt-file (CB-618 forbids the inline flag and the file flag together),
+        // always — fleetd #220: charter prose on the command line overruns the pane's byte cap.
         List<String> charters = new java.util.ArrayList<>(2);
         if (roleCharter != null) charters.add(roleCharter);
         if (replyCharter != null) charters.add(replyCharter);
-        if (charters.size() == 1 && replyCharter != null && roleCharter == null) {
-            argv.add("--append-system-prompt");
-            argv.add(replyCharter);
-        } else if (!charters.isEmpty()) {
+        if (!charters.isEmpty()) {
             argv.add("--append-system-prompt-file");
             argv.add(writeCharterFile(String.join("\n\n", charters)).toString());
         }
@@ -450,12 +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/EnvAllowListScrub.java

@@ -143,17 +143,24 @@ public final class EnvAllowListScrub {
     }
 
     /**
-     * chgrp/chmod-equivalent over the freshly generated directory and the startup files already
-     * written into it: owner keeps full access, {@code group} gets traverse+read on the directory
-     * ({@code rwxr-x---}, so a login shell under that group can find and source the files) and
-     * read-only on each file ({@code rw-r-----}) — deliberately no group WRITE anywhere, since a
-     * member never needs to add or change fleetd's own generated scrub. (The scrub script's own
-     * report write inside the pane consequently fails closed rather than open — see {@code
-     * scrub.zsh}'s trailing {@code 2>/dev/null} — which {@link
+     * chgrp/chmod-equivalent over a freshly generated directory and the flat files already written
+     * into it: owner keeps full access, {@code group} gets traverse+read on the directory ({@code
+     * rwxr-x---}, so a member process — a login shell reading it via {@code ZDOTDIR}, or another
+     * process simply opening a file under it — running under that group can find and read the
+     * files) and read-only on each file ({@code rw-r-----}) — deliberately no group WRITE anywhere,
+     * since a member never needs to add or change what fleetd generated. (For the ZDOTDIR scrub
+     * specifically, this also means the scrub script's own report write inside the pane fails
+     * closed rather than open — see {@code scrub.zsh}'s trailing {@code 2>/dev/null} — which {@link
      * dev.ltms.fleet.member.HerdrPeerLauncher#releaseZdotdir} already treats as "cannot be
      * confirmed to have run" rather than success.)
+     *
+     * <p>Package-private and named generically on purpose: fleetd #213 built this for the ZDOTDIR
+     * scrub directory, and fleetd #219 reuses it verbatim for {@link
+     * dev.ltms.fleet.member.OpenCodeLauncher}'s ephemeral {@code opencode.json} directory — both are
+     * "a fleetd-generated directory of flat files that a different-uid member process must read but
+     * never write," so the sharing mechanism is shared rather than copied a second time.
      */
-    private static void shareWithGroup(Path dir, String group) {
+    static void shareWithGroup(Path dir, String group) {
         try {
             GroupPrincipal principal = dir.getFileSystem().getUserPrincipalLookupService()
                     .lookupPrincipalByGroupName(group);
@@ -164,11 +171,11 @@ public final class EnvAllowListScrub {
                 }
             }
         } catch (IOException e) {
-            throw new UncheckedIOException("cannot share generated ZDOTDIR " + dir + " with group '"
+            throw new UncheckedIOException("cannot share generated directory " + dir + " with group '"
                     + group + "' — the group must exist, and the fleetd operator ("
                     + System.getProperty("user.name") + ") must be a member of it", e);
         } catch (UnsupportedOperationException e) {
-            throw new UncheckedIOException("cannot share generated ZDOTDIR " + dir + " with group '"
+            throw new UncheckedIOException("cannot share generated directory " + dir + " with group '"
                     + group + "' — this filesystem does not support POSIX group ownership",
                     new IOException(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 -------------------------------------------------------------------------
@@ -682,6 +693,7 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         // Protocol 19 resolves the executable from the agent kind (== namePrefix here), so
         // argv[0] — the configured executable — is dropped and only the extra args are passed.
         List<String> args = argv.isEmpty() ? argv : argv.subList(1, argv.size());
+        checkPaneCommandFits(cfg, argv);
         HerdrException last = null;
         for (int attempt = 0; attempt < NAME_RETRIES; attempt++) {
             long seq = nameSeq.incrementAndGet();
@@ -697,6 +709,59 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         throw last;
     }
 
+    /**
+     * fleetd #220: herdr does not exec the launch command — it TYPES it into the pane as one line,
+     * and a pty line buffer holds only {@value #PANE_COMMAND_BYTE_LIMIT} bytes (BSD/macOS {@code
+     * MAX_CANON}). Everything past that byte is dropped. Nothing reports it: herdr answers "agent
+     * started", the backend exits on the mangled argument it was handed, the pane closes, and the
+     * only symptom is {@link #waitUntilInjectableOrThrow} timing out 20 seconds later with no
+     * reason. That is exactly how #214 broke every claude-code spawn — one 50-byte flag pushed a
+     * 978-byte command to 1028, and the tail that got cut was {@code --autocompact 250000}.
+     *
+     * <p>So measure it here and refuse, loudly and immediately, rather than spawn something that
+     * cannot work. The estimate is deliberately conservative: fleetd cannot see herdr's quoting, so
+     * every argument is charged its own bytes plus a separator and a quote pair. An over-estimate
+     * costs a clear error at a length that was already unsafe; an under-estimate would let the
+     * silent truncation back in.
+     *
+     * @throws PeerUnreachableException when the command cannot fit — the same failure the spawn
+     *                                  would have hit anyway, named at the point it is still
+     *                                  explainable
+     */
+    private void checkPaneCommandFits(FleetConfig.Profile cfg, List<String> argv) {
+        int bytes = 0;
+        String longest = null;
+        int longestBytes = 0;
+        for (String arg : argv) {
+            int argBytes = arg == null ? 0 : arg.getBytes(java.nio.charset.StandardCharsets.UTF_8).length;
+            bytes += argBytes + QUOTING_OVERHEAD_PER_ARG;
+            if (argBytes > longestBytes) {
+                longestBytes = argBytes;
+                longest = arg;
+            }
+        }
+        if (bytes <= PANE_COMMAND_BYTE_LIMIT) {
+            return;
+        }
+        String culprit = longest == null ? "<none>"
+                : longest.substring(0, Math.min(longest.length(), 60)) + (longest.length() > 60 ? "…" : "");
+        throw new PeerUnreachableException(
+                "launch command for profile " + cfg.profile() + " is about " + bytes + " bytes, over the "
+                        + PANE_COMMAND_BYTE_LIMIT + "-byte limit of the pane line herdr types it into. "
+                        + "The pty would drop the tail silently and the backend would exit on a mangled "
+                        + "argument. Longest argument is " + longestBytes + " bytes: " + culprit
+                        + " — move it off the command line (a file flag) or shorten it.");
+    }
+
+    /**
+     * The pty line buffer herdr types a launch command into: BSD/macOS {@code MAX_CANON}. Not a
+     * fleetd choice and not configurable — see {@link #checkPaneCommandFits}.
+     */
+    static final int PANE_COMMAND_BYTE_LIMIT = 1024;
+
+    /** Per-argument allowance for the separating space and a shell quote pair fleetd cannot see. */
+    private static final int QUOTING_OVERHEAD_PER_ARG = 3;
+
     /** Start the agent into {@code paneId}, waiting out the seed shell's boot with the sleeper. */
     private Agent startAwaitingShellPrompt(String name, List<String> args, String paneId) {
         HerdrException busy = null;
@@ -855,25 +920,135 @@ 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;
+        long start = nowMillis.getAsLong();
+        long deadline = start + spawnReadyTimeoutMs;
+        AgentStatus lastStatus = null;
         while (nowMillis.getAsLong() < deadline) {
-            if (agents.status(paneId).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;
             }
             sleeper.run();
         }
-        log.warn("peer pane={} did not become injectable within {}ms — closing", paneId, spawnReadyTimeoutMs);
+        // fleetd #220: read the pane BEFORE stop() closes it. Without this the gate says only that
+        // it timed out, which is true of every cause — a backend that never launched, a binary that
+        // rejected an argument and exited, a trust prompt, a login shell that hung. The pane holds
+        // the one copy of that answer and it is destroyed a line later.
+        log.warn("peer pane={} did not become injectable within {}ms (last status {}) — closing. "
+                        + "Pane tail:\n{}",
+                paneId, spawnReadyTimeoutMs, lastStatus, readPaneQuietly(paneId));
         stop(paneId);
         throw new PeerUnreachableException(
                 "worker pane " + paneId + " did not reach injectable state within "
                         + 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
+     * already failing, so it must never replace the real error with one of its own.
+     */
+    private String readPaneQuietly(String paneId) {
+        try {
+            String pane = agents.read(paneId, "recent");
+            if (pane == null || pane.isBlank()) {
+                return "<pane read returned nothing>";
+            }
+            return pane.length() <= SPAWN_FAILURE_PANE_CHARS
+                    ? pane
+                    : pane.substring(pane.length() - SPAWN_FAILURE_PANE_CHARS);
+        } catch (RuntimeException e) {
+            return "<pane could not be read: " + e.getMessage() + ">";
+        }
+    }
+
+    /** How much of a failed spawn's pane the timeout log carries. */
+    private static final int SPAWN_FAILURE_PANE_CHARS = 4000;
+
     /**
      * A concrete {@link PeerHandle} wrapping herdr agent coordinates, the profile that spawned it,
      * the session identity the launch resolved (CB-547a): the bridge's logical name and the peer's
@@ -1190,8 +1365,13 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * provisioning a worktree that will exist regardless, whereas an unconfigured value here means
      * fleetd has no operator-endorsed location to put a credential-bearing directory a different OS
      * user must reach, so falling back to the overlay is the honest answer, not a guess.
+     *
+     * <p>Package-private (fleetd #219) so {@link OpenCodeLauncher} can reuse the exact same
+     * "different OS user, put it under worktreeRoot instead of java.io.tmpdir" resolution for its
+     * own ephemeral {@code opencode.json} directory, rather than re-reading {@code config} a second
+     * time with a second copy of this null/blank handling.
      */
-    private Path memberScrubParentDir() {
+    Path memberScrubParentDir() {
         FleetConfig cfg = config == null ? null : config.get();
         if (cfg == null || cfg.worktreeRoot() == null || cfg.worktreeRoot().isBlank()) {
             return null;
@@ -1204,8 +1384,11 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * the group {@link dev.ltms.fleet.session.Worktrees#shareWithGroup} already establishes for
      * provisioned worktrees, rather than a second group key — see {@link
      * #applyEnvironmentAllowListPolicy}.
+     *
+     * <p>Package-private (fleetd #219) — reused by {@link OpenCodeLauncher} alongside {@link
+     * #memberScrubParentDir()}; see that method's javadoc.
      */
-    private String memberGroup() {
+    String memberGroup() {
         FleetConfig cfg = config == null ? null : config.get();
         if (cfg == null || cfg.worktreeGroup() == null || cfg.worktreeGroup().isBlank()) {
             return null;
@@ -1378,8 +1561,12 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * through (every production {@code HerdrPeerLauncher} does; a handful of older tests do not) —
      * treated the same as "not configured", which is the correct, permissive default: it is exactly
      * today's single-daemon behaviour.
+     *
+     * <p>Package-private (fleetd #219) — {@link OpenCodeLauncher} reuses this same gate to decide
+     * where its own ephemeral {@code opencode.json} directory (site 1) and its opencode session
+     * discovery (site 2) may run, rather than re-deriving "is this a multi-uid fleet" a second way.
      */
-    private boolean memberHerdrSocketConfigured() {
+    boolean memberHerdrSocketConfigured() {
         if (config == null) {
             return false;
         }

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;
@@ -20,6 +21,8 @@ import java.util.EnumSet;
 import java.util.List;
 import java.util.Map;
 import java.util.Set;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.function.BooleanSupplier;
 import java.util.function.Function;
 import java.util.function.LongSupplier;
 import java.util.function.Supplier;
@@ -71,6 +74,19 @@ 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.
@@ -130,9 +146,25 @@ 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);
     }
 
     /**
@@ -180,6 +212,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                 spawnReadyTimeoutMs, nowMillis, sleeper, fleet);
         this.configRoot = configRoot;
         this.discovery = new OpenCodeSessionDiscovery(discoveryRoot);
+        this.exhaustionSink = ExhaustionSink.none();
     }
 
     /**
@@ -205,17 +238,72 @@ 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() {
         return Path.of(System.getProperty("java.io.tmpdir"));
     }
 
-    /** The default opencode storage root: {@code ~/.local/share/opencode} (the XDG data dir). */
+    /**
+     * The default opencode storage root: {@code ~/.local/share/opencode} (the XDG data dir) —
+     * always FLEETD's OWN {@code user.home}, whichever OS user runs the daemon.
+     *
+     * <p><b>fleetd #219 site 2 — a decision, not a patch.</b> Under {@code memberHerdrSocket:} the
+     * member pane runs as a <em>different</em> OS user, and opencode writes {@code opencode.db}
+     * under <em>that</em> user's {@code $HOME}, not fleetd's. Scanning fleetd's own {@code
+     * user.home} is therefore looking in the wrong place — a wrong-LOCATION failure, not a
+     * wrong-PERMISSION one like site 1, and it fails quietly: {@link
+     * SessionAwareHandle#agentSessionId()} would keep returning {@code null} forever, which reads
+     * as "opencode does not support resume" rather than "fleetd looked in the wrong home." fleetd
+     * #209 is the reason that silence is unacceptable.
+     *
+     * <p>Three ways to close the gap were weighed:
+     * <ol>
+     *   <li><b>Make the member's home configurable.</b> Correct in principle, but this ticket's
+     *       scope is the two existing call sites, not a new config key — {@code memberHerdrSocket}
+     *       already carries the second herdr's socket path, not its user's home, and inventing a
+     *       parallel key here without also wiring it through discovery's actual callers is a
+     *       half-shipped feature (the exact shape CB-596/CB-611 warn against).</li>
+     *   <li><b>Derive it</b> (e.g. from {@code worktreeRoot}'s owner, or {@code getent passwd}).
+     *       Rejected: nothing in this codebase resolves a Unix username to a home directory today,
+     *       and guessing wrong would silently point discovery at a THIRD wrong location — worse
+     *       than the current gap, because it would look like it should work.</li>
+     *   <li><b>Declare discovery unavailable</b> under {@code memberHerdrSocket}, and say so once,
+     *       loudly, instead of scanning a directory that structurally cannot hold the answer.</li>
+     * </ol>
+     *
+     * <p>Option 3 is taken — the one this ticket says to default to when unsure. {@link
+     * OpenCodeLauncher#spawn} routes {@link SessionAwareHandle#agentSessionId()} through {@link
+     * HerdrPeerLauncher#memberHerdrSocketConfigured()} before ever calling {@link
+     * OpenCodeSessionDiscovery#sessionIdForDirectory}, so under {@code memberHerdrSocket} the
+     * database at this root is never even opened, and one WARN per launcher instance names the gap
+     * instead of the {@code null} return reading as "unsupported." Capability advertising is
+     * unaffected: {@link #capabilities()} always includes {@code SESSION_RESUME}, since {@code
+     * memberHerdrSocket} absent (today's only live mode) is unchanged by this decision.
+     */
     private static Path defaultDiscoveryRoot() {
         return Path.of(System.getProperty("user.home"), ".local", "share", "opencode");
     }
@@ -331,7 +419,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
      */
     private Path writeConfig(FleetConfig.Profile cfg, String charterText, String cwd) {
         try {
-            Path dir = Files.createTempDirectory(configRoot, "fleetd-opencode-");
+            Path dir = Files.createTempDirectory(configParentDir(), "fleetd-opencode-");
             dir.toFile().deleteOnExit();
 
             ObjectNode root = JSON.createObjectNode();
@@ -402,6 +490,14 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
             // carries operator-supplied values (URL, model id, api key), so escaping must be real.
             Files.writeString(cfgFile, JSON.writerWithDefaultPrettyPrinter().writeValueAsString(root));
             cfgFile.toFile().deleteOnExit();
+            if (memberHerdrSocketConfigured()) {
+                // fleetd #219: the same "different OS user" gap fleetd #213 closed for the ZDOTDIR
+                // scrub — share read-only with worktreeGroup rather than leaving the directory under
+                // fleetd's own 0700 java.io.tmpdir, where the member's OS user could not even
+                // traverse it. memberGroup() cannot be null here: configParentDir() above already
+                // refused this spawn if either worktreeRoot or worktreeGroup was missing.
+                EnvAllowListScrub.shareWithGroup(dir, memberGroup());
+            }
             return cfgFile;
         } catch (IOException e) {
             throw new UncheckedIOException(
@@ -409,6 +505,57 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
         }
     }
 
+    /**
+     * fleetd #219 site 1: where {@link #writeConfig} creates its per-spawn directory.
+     *
+     * <ul>
+     *   <li>{@code memberHerdrSocket} ABSENT (today's only mode): byte-identical to before this
+     *       fix — always {@link #configRoot} (defaults to {@code java.io.tmpdir}, fleetd's own
+     *       process).</li>
+     *   <li>{@code memberHerdrSocket} PRESENT: {@code java.io.tmpdir} is fleetd's own per-user temp
+     *       dir (mode {@code 0700} on macOS) — the member pane runs as a DIFFERENT OS user under
+     *       this config key and cannot even traverse it, so the directory holding {@code
+     *       opencode.json} (which tells the member where the bridge MCP is) and the member charter
+     *       would be unreadable to the very process it is written for. The directory instead goes
+     *       under {@code worktreeRoot}, shared read-only with {@code worktreeGroup} via {@link
+     *       EnvAllowListScrub#shareWithGroup} — the SAME mechanism fleetd #213 built for the ZDOTDIR
+     *       scrub, reused here rather than duplicated (see {@link
+     *       HerdrPeerLauncher#memberScrubParentDir()}).</li>
+     * </ul>
+     *
+     * <p><b>Unlike the ZDOTDIR scrub, a missing {@code worktreeRoot}/{@code worktreeGroup} here
+     * REFUSES the spawn instead of degrading.</b> The ZDOTDIR scrub is a credential CONTROL: a
+     * degraded control (CB-596's sentinel overlay) is still worth having. This config file is not a
+     * control — it is the ONLY way the member learns where the bridge MCP lives. Writing it
+     * somewhere the member cannot read would not degrade anything; it would spawn a member that
+     * occupies a pane and never becomes deliverable, since {@code fleet_send} waits ~60s on the
+     * readiness gate and then fails with nothing pointing at a temp directory as the cause. Refusing
+     * up front, with a message that names the missing config key, is the honest failure — an
+     * undeliverable member is not a working spawn either way, so nothing is lost by refusing loudly
+     * instead of failing silently later.
+     *
+     * @throws IllegalStateException when {@code memberHerdrSocket} is configured but {@code
+     *                                worktreeRoot} and/or {@code worktreeGroup} is not
+     */
+    private Path configParentDir() {
+        if (!memberHerdrSocketConfigured()) {
+            return configRoot;
+        }
+        Path root = memberScrubParentDir();
+        String group = memberGroup();
+        if (root == null || group == null) {
+            throw new IllegalStateException("memberHerdrSocket is configured, so opencode's config "
+                    + "directory (opencode.json + member charter) must be placed where the member's "
+                    + "OS user can read it — worktreeRoot, shared via worktreeGroup — but "
+                    + (root == null ? "worktreeRoot" : "worktreeGroup") + " is not configured. "
+                    + "Refusing to spawn rather than write a config the member cannot read: that "
+                    + "member would occupy a pane and never become deliverable, with nothing "
+                    + "pointing at the real cause. Configure both worktreeRoot and worktreeGroup to "
+                    + "enable opencode member spawns under memberHerdrSocket.");
+        }
+        return root;
+    }
+
     /**
      * Declare a custom OpenAI-compatible provider so the worker talks to a pinned endpoint (a local
      * vLLM, say) instead of opencode's default gateway (CB-508).
@@ -517,11 +664,21 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
         return afterScheme.contains("/") ? trimmed : trimmed + "/v1";
     }
 
+    /** One WARN per launcher instance for the fleetd #219 site-2 discovery-unavailable gap. */
+    private final AtomicBoolean discoveryUnavailableWarned =
+            new AtomicBoolean();
+
     /** Add lazy on-disk session discovery to the base handle. */
     @Override
     public PeerHandle spawn(SpawnRequest req) {
         PeerHandle inner = super.spawn(req);
-        return new SessionAwareHandle(inner, discovery, effectiveCwd(req));
+        // 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);
     }
 
     /**
@@ -531,16 +688,37 @@ 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();
 
-        SessionAwareHandle(PeerHandle delegate, OpenCodeSessionDiscovery discovery, String cwd) {
+        SessionAwareHandle(PeerHandle delegate, OpenCodeSessionDiscovery discovery, String cwd,
+                            FleetConfig.Profile cfg,
+                            BooleanSupplier discoveryUnavailable,
+                            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
@@ -565,11 +743,84 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
 
         @Override
         public String agentSessionId() {
+            // fleetd #219 site 2: under memberHerdrSocket the member pane runs as a different OS
+            // user, so opencode.db lives under THAT user's $HOME, not the one discoveryRoot was
+            // built from (see OpenCodeLauncher#defaultDiscoveryRoot's javadoc for the full
+            // reasoning). Scanning fleetd's own $HOME under that config would only ever find "no
+            // row" and read as "resume unsupported" — declare it unavailable instead, once, loudly.
+            if (discoveryUnavailable.getAsBoolean()) {
+                if (discoveryUnavailableWarned.compareAndSet(false, true)) {
+                    log.warn("opencode session discovery unavailable: memberHerdrSocket is "
+                            + "configured, so opencode's on-disk session database lives under the "
+                            + "MEMBER's own $HOME, not fleetd's ({}) — agentSessionId will stay null "
+                            + "for every opencode member under this config, and SESSION_RESUME "
+                            + "cannot be honored (fleetd #209/#219).",
+                            System.getProperty("user.home"));
+                }
+                return null;
+            }
             // 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);
+            // 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();
+            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, 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.
+         */
+        private void checkModelMatch() {
+            if (modelMismatchReported.get() || cfg.model() == null || cfg.model().isBlank()) {
+                return;
+            }
+            OpenCodeSessionDiscovery.ActualModel actual = discovery.actualModelForDirectory(cwd);
+            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);
+            exhaustionSink.onExhausted(delegate.terminalId(),
+                    "opencode model mismatch: profile '" + cfg.profile() + "' requested '"
+                            + cfg.model() + "' but the live session is running '" + actualDisplay
+                            + "' (fleetd #175)");
         }
 
         @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;
@@ -44,6 +47,18 @@ 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;
@@ -117,4 +132,73 @@ final class OpenCodeSessionDiscovery {
                 storageRoot, directory);
         return null;
     }
+
+    /**
+     * The model opencode actually ran the {@code directory}'s most-recent session on (fleetd
+     * #175), read from the same row {@link #sessionIdForDirectory} matches — but via its own
+     * query and its own connection, deliberately kept independent so a database whose schema
+     * predates the {@code model} column (or any other read failure on this column alone) can
+     * never take {@link #sessionIdForDirectory}'s 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.
+     *
+     * @param directory the worker's cwd, as resolved for this spawn
+     * @return the actual model, or {@code null} when unknown
+     */
+    ActualModel actualModelForDirectory(String directory) {
+        if (directory == null || directory.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 directory = ? ORDER BY time_updated DESC LIMIT 1";
+        try (Connection connection = openReadOnly();
+             PreparedStatement statement = connection.prepareStatement(sql)) {
+            statement.setString(1, directory);
+            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

@@ -4,8 +4,11 @@ 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 com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
 import dev.ltms.fleet.herdr.AgentControl;
 import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
 import dev.ltms.fleet.msg.Rendezvous;
 import dev.ltms.fleet.msg.TestTurnTokens;
 import dev.ltms.fleet.msg.TurnToken;
@@ -633,6 +636,115 @@ class CompletionResolverTest {
                 "the failure carries the rest of the pane, not only the matched line: " + reason);
     }
 
+    // --- fleetd#211: raw-scrape fallback classification when there is no usable assistant block ---
+
+    @Test
+    void anExhaustionLineWithNoMarkerAndLeadingChromeIsClassifiedFromTheRawScrapeAndNotifiesTheSink() {
+        // No ⏺ anywhere, and the first visible line is TUI chrome (╭). lastAssistantBlock's boundary
+        // scan starts at the top of the raw screen and breaks immediately, so the trimmed block is "".
+        // The fix: fall back to matching the RAW scrape so this doesn't get lost as an empty scrape.
+        String block = """
+                ╭──────────────────────────────────────╮
+                The usage limit has been reached. Try again later.
+                """;
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        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);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertTrue(waiter.isDone(), "a raw-scrape match still resolves the blocked send");
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "classified from the raw scrape even though the trimmed block was empty");
+        assertEquals(1, notified.size(),
+                "the sink is the whole point of this ticket — it must be notified: " + notified);
+        assertTrue(notified.get(0).startsWith("term_a: "), "the sink is told which target exhausted");
+        assertTrue(notified.get(0).contains("The usage limit has been reached"),
+                "the sink is told the matched reason: " + notified.get(0));
+    }
+
+    @Test
+    void aBackendErrorLineWithNoMarkerAndLeadingChromeIsClassifiedFromTheRawScrape() {
+        String block = """
+                ╭──────────────────────────────────────╮
+                API Error: 400 invalid request body
+                """;
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, ExhaustedPatternLookup.none(), ExhaustionSink.none());
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertTrue(waiter.isDone(), "a raw-scrape backend-error match still resolves the blocked send");
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind(),
+                "classified BACKEND_ERROR from the raw scrape even though the trimmed block was empty");
+        assertTrue(waiter.getNow(null).text().contains("API Error: 400 invalid request body"),
+                "the failure carries the matched line: " + waiter.getNow(null).text());
+        assertTrue(waiter.getNow(null).text().contains("--- pane tail ---"),
+                "fleetd#164: the failure must carry the pane, not only the matched line — with an "
+                        + "empty trimmed block the raw scrape is the only copy of what the member said: "
+                        + waiter.getNow(null).text());
+        assertTrue(waiter.getNow(null).text().contains("╭"),
+                "the carried pane is the raw scrape, chrome included: " + waiter.getNow(null).text());
+    }
+
+    @Test
+    void anOrdinaryPaneWithANormalAssistantBlockIsUnaffectedByTheRawScrapeFallback() {
+        // Pin: on a pane that already yields a usable block, the fallback branch is never reached —
+        // same outcome, same text, sink not called — even though the raw screen around the marker
+        // would itself match the configured exhausted pattern.
+        String block = "The usage limit has been reached, but this is a leading TUI line above the "
+                + "marker.\n⏺ complete report\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        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);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertTrue(waiter.isDone());
+        assertEquals(Rendezvous.Kind.COMPLETION, waiter.getNow(null).kind(),
+                "unchanged: a usable assistant block never reaches the raw-scrape fallback");
+        assertEquals("complete report", waiter.getNow(null).text(), "the reply text is unaffected");
+        assertTrue(notified.isEmpty(), "the fallback never runs, so the sink is never called");
+    }
+
+    @Test
+    void aGenuinelyEmptyScrapeStillFailsAsEmptyAndNeverNotifiesTheSink() {
+        // The false-positive pin: no exhaustion or backend-error text anywhere on the pane (just
+        // chrome, no marker) — the raw-scrape fallback must not manufacture a classification, and
+        // the sink must stay untouched.
+        String block = """
+                ╭──────────────────────────────────────╮
+                │ >                                    │
+                ╰──────────────────────────────────────╯
+                """;
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        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);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous, patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertTrue(waiter.isDone(), "an empty scrape must still resolve the send, not hang");
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind(),
+                "no exhaustion or backend-error text anywhere ⇒ this stays the ordinary empty-scrape failure");
+        assertTrue(waiter.getNow(null).text().toLowerCase().contains("empty"),
+                "the failure still says the scrape was empty: " + waiter.getNow(null).text());
+        assertTrue(notified.isEmpty(), "a genuinely empty pane must never quarantine a credential");
+    }
+
     @Test
     void coverageIsOffWhenNoProfileHasAPatternConfigured() {
         assertEquals("off (no profile has an exhaustedPattern configured; profiles: [terra])",
@@ -650,4 +762,202 @@ class CompletionResolverTest {
         assertEquals("partial (configured: [terra]; not configured: [gx10])",
                 CompletionResolver.coverage(Set.of("terra", "gx10"), Set.of("terra")));
     }
+
+    // --- fleetd#201 Unit 1: target-keyed backend-error pattern + typed sink ----------------------
+
+    @Test
+    void aConfiguredBackendErrorPatternClassifiesAMatchAsAFailureAndNotifiesTheSinkOnce() {
+        String block = "⏺ 503 Service Unavailable: upstream credential rejected\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertTrue(waiter.isDone(), "a configured-pattern match still resolves the blocked send");
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind());
+        assertEquals(1, notified.size(), "the sink fires exactly once for the winning classification");
+        assertEquals("term_a: 503 Service Unavailable: upstream credential rejected", notified.get(0));
+    }
+
+    @Test
+    void aLosingBackendErrorClassificationNeverNotifiesTheSink() {
+        // The waiter was already resolved (e.g. by the worker's own fleet_reply) before this scrape
+        // landed — resolveFailure loses the race and must return false, so the sink must not fire.
+        String block = "⏺ 503 Service Unavailable: upstream credential rejected\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        var turn = new CompletionResolver.InFlight(waiter, null);
+        assertTrue(rendezvous.resolveCompletion(waiter, "already replied")); // fleet_reply won first
+
+        resolver.resolve("term_a", turn);
+
+        assertTrue(notified.isEmpty(), "a classification that loses the race must not fire the sink");
+        assertEquals("already replied", waiter.getNow(null).text(), "the earlier resolution stands untouched");
+    }
+
+    @Test
+    void aBackendErrorClassificationThatLosesARealRaceDuringTheScrapeNeverNotifiesTheSink() {
+        // The test above pre-resolves the waiter BEFORE calling resolve(), so it is caught by
+        // resolve()'s own top-of-method isDone() guard before ever reaching the win-gated
+        // resolveFailure call this unit adds — it proves the outcome, not the specific gate. This
+        // test forces the actual race window fleetd#201 must be safe in: a genuine fleet_reply lands
+        // WHILE the herdr scrape for this turn is in flight, i.e. AFTER resolve() has already passed
+        // the top guard (waiter not done yet) and committed to reading the pane, but BEFORE it
+        // reaches the "if (rendezvous.resolveFailure(...))" line below. The side effect is attached
+        // to the one real I/O step resolve() performs between those two points: the herdr
+        // "agent.read" call itself.
+        Rendezvous rendezvous = new Rendezvous();
+        var waiter = rendezvous.open("term_a");
+        ObjectMapper mapper = new ObjectMapper();
+        HerdrClient racingDuringScrape = new HerdrClient() {
+            @Override
+            public JsonNode call(String method, Object params) {
+                if ("agent.read".equals(method)) {
+                    // The real fleet_reply that wins the race, landing mid-scrape.
+                    assertTrue(rendezvous.resolve("term_a", "a genuine fleet_reply landed first"),
+                            "the racing reply must land while the waiter is still open");
+                }
+                try {
+                    return mapper.readTree(mapper.writeValueAsString(java.util.Map.of(
+                            "type", "agent_read",
+                            "read", java.util.Map.of("text",
+                                    "⏺ 503 Service Unavailable: upstream credential rejected\n❯ "))));
+                } catch (Exception e) {
+                    throw new RuntimeException(e);
+                }
+            }
+
+            @Override
+            public void close() {
+            }
+        };
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(racingDuringScrape), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink);
+
+        var turn = new CompletionResolver.InFlight(waiter, null); // not done yet — passes the top guard
+        resolver.resolve("term_a", turn);
+
+        assertTrue(notified.isEmpty(),
+                "a classification that loses a real race during the scrape must not fire the sink");
+        assertEquals(Rendezvous.Kind.REPLY, waiter.getNow(null).kind(), "the genuine fleet_reply stands");
+        assertEquals("a genuine fleet_reply landed first", waiter.getNow(null).text());
+    }
+
+    @Test
+    void exhaustionKeepsWinningOverBackendErrorEvenWhenBothPatternsMatchTheSameLine() {
+        // A line matching both a configured exhausted pattern AND a configured backend-error pattern
+        // must classify BACKEND_EXHAUSTED and call only the ExhaustionSink — the ordering fleetd#578
+        // already relies on must not change.
+        String block = "⏺ The usage limit has been reached. Try again later.\n❯ ";
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        ExhaustedPatternLookup exhausted = target -> Pattern.compile("usage limit has been reached");
+        BackendErrorPatternLookup backendErrors = target -> Pattern.compile("(?i)usage limit");
+        java.util.List<String> exhaustedNotified = new java.util.ArrayList<>();
+        java.util.List<String> backendErrorNotified = new java.util.ArrayList<>();
+        ExhaustionSink exhaustionSink = (target, reason) -> exhaustedNotified.add(target);
+        BackendErrorSink backendErrorSink = (target, matchedLine, reason) -> backendErrorNotified.add(target);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                exhausted, exhaustionSink, backendErrors, backendErrorSink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertEquals(Rendezvous.Kind.BACKEND_EXHAUSTED, waiter.getNow(null).kind(),
+                "exhaustion must keep winning over backend error");
+        assertEquals(1, exhaustedNotified.size(), "the exhaustion sink fires");
+        assertTrue(backendErrorNotified.isEmpty(), "the backend-error sink must never fire for this turn");
+    }
+
+    @Test
+    void aConfiguredPatternAlsoClassifiesTheRawScrapeFallbackAndNotifiesTheSink() {
+        // No ⏺ marker and leading TUI chrome ⇒ lastAssistantBlock() yields "", so classification must
+        // fall back to the raw scrape (fleetd#211) — and it must use the configured pattern too.
+        String block = """
+                ╭──────────────────────────────────────╮
+                503 Service Unavailable: upstream credential rejected
+                """;
+        FakeHerdr herdr = new FakeHerdr().readText(block);
+        Rendezvous rendezvous = new Rendezvous();
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink);
+
+        var waiter = rendezvous.open("term_a");
+        resolver.resolve("term_a", new CompletionResolver.InFlight(waiter, null));
+
+        assertTrue(waiter.isDone(), "a raw-scrape configured-pattern match still resolves the blocked send");
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind(),
+                "classified from the raw scrape even though the trimmed block was empty");
+        assertEquals(1, notified.size(), "the sink fires exactly once from the raw-scrape path");
+        assertTrue(notified.get(0).contains("503 Service Unavailable"), notified.get(0));
+    }
+
+    // --- fleetd#201 Unit 1: classification inside the fleetd#164 MIN_TURN_NANOS floor -------------
+
+    @Test
+    void aMatchingErrorInsideTheFloorIsTypedAndNotifiesTheSink() {
+        FakeHerdr herdr = new FakeHerdr().readText("⏺ 503 Service Unavailable: upstream credential rejected\n❯ ");
+        Rendezvous rendezvous = new Rendezvous();
+        long[] clock = {10_000_000_000L};
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink, () -> clock[0]);
+
+        var waiter = rendezvous.open("term_a");
+        var turn = new CompletionResolver.InFlight(waiter, null, clock[0]); // delivered "now"
+        clock[0] += CompletionResolver.MIN_TURN_NANOS - 1; // inside the floor
+
+        resolver.resolve("term_a", turn);
+
+        assertTrue(waiter.isDone());
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind());
+        assertEquals(1, notified.size(), "a matching error inside the floor notifies the sink");
+        assertTrue(notified.get(0).contains("503 Service Unavailable"), notified.get(0));
+    }
+
+    @Test
+    void aNonMatchInsideTheFloorStaysGenericAndNeverNotifiesTheSink() {
+        FakeHerdr herdr = new FakeHerdr().readText("⏺ still starting up\n❯ ");
+        Rendezvous rendezvous = new Rendezvous();
+        long[] clock = {10_000_000_000L};
+        BackendErrorPatternLookup patterns = target -> Pattern.compile("(?i)503 Service Unavailable");
+        java.util.List<String> notified = new java.util.ArrayList<>();
+        BackendErrorSink sink = (target, matchedLine, reason) -> notified.add(target + ": " + matchedLine);
+        CompletionResolver resolver = new CompletionResolver(new AgentControl(herdr), rendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(), patterns, sink, () -> clock[0]);
+
+        var waiter = rendezvous.open("term_a");
+        var turn = new CompletionResolver.InFlight(waiter, null, clock[0]); // delivered "now"
+        clock[0] += CompletionResolver.MIN_TURN_NANOS - 1; // inside the floor
+
+        resolver.resolve("term_a", turn);
+
+        assertTrue(waiter.isDone());
+        assertEquals(Rendezvous.Kind.FAILED, waiter.getNow(null).kind(),
+                "still a failure — the floor itself, not the pattern, is why");
+        assertTrue(waiter.getNow(null).text().contains("too fast to be real work"),
+                "a non-match inside the floor stays the generic too-fast reason: " + waiter.getNow(null).text());
+        assertTrue(notified.isEmpty(), "a non-match must never notify the typed sink");
+    }
 }

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 {
@@ -61,8 +64,75 @@ class ClaudeCodeLauncherTest {
         assertTrue(args.stream().noneMatch(a -> a.contains("\"bridge\"")),
                 "the mount is named fleet since CB-632 — a member addresses its tools as "
                         + "mcp__fleet__*, and CLAUDE.md's role-detection ladder names that prefix");
-        assertTrue(args.contains("--append-system-prompt"));
-        assertTrue(args.stream().anyMatch(a -> a.contains("fleet_reply")), "reply charter present");
+        // fleetd #220: the charter travels as a FILE, never inline — charter prose on the command
+        // line overruns the byte cap of the pane line herdr types it into.
+        assertTrue(args.contains("--append-system-prompt-file"));
+        assertFalse(args.contains("--append-system-prompt"),
+                "the inline flag would put ~800 bytes of prose on the pane command line");
+        String charterFile = args.get(args.indexOf("--append-system-prompt-file") + 1);
+        assertTrue(readFile(charterFile).contains("fleet_reply"), "reply charter present in the file");
+    }
+
+    /** Read a charter file the launcher wrote, failing the test rather than the build on an IO error. */
+    private static String readFile(String path) {
+        try {
+            return java.nio.file.Files.readString(java.nio.file.Path.of(path));
+        } catch (java.io.IOException e) {
+            throw new AssertionError("charter file " + path + " is not readable", e);
+        }
+    }
+
+    /**
+     * fleetd #220 regression. herdr TYPES the launch command into the pane, and the pty line buffer
+     * holds 1024 bytes — past that the tail is dropped with no error anywhere, so the backend exits
+     * on a mangled argument and the spawn dies as an unexplained readiness timeout. That is what
+     * happened when #214 added --session-id to a command already 978 bytes long: --autocompact
+     * 250000 arrived as --autocompact 25. This asserts the whole assembled command still fits, with
+     * the flags a real spawn carries (MCP mount, charter, model, autocompact, session id).
+     */
+    @Test
+    void theAssembledLaunchCommandFitsThePaneLineLimit() {
+        FakeHerdr herdr = new FakeHerdr();
+        // Shaped like the live sonnet profile, because the bug is a SUM: the charter alone fits,
+        // and so does every flag alone. Only model + autocompact + session id on top of the charter
+        // crossed the cap, which is why nothing caught it until a member failed to spawn.
+        FleetConfig.Profile cfg = new FleetConfig.Profile(
+                "sonnet", null, "claude-sonnet-5", null, null,
+                List.of("claude"), "tab", "fleetd-workers", "w #{n}", "http://127.0.0.1:8765/mcp",
+                null, null, null, null, null, null, null, null, true, null, null, null, null, null,
+                250000);
+        new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of()), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null)
+                .spawn();
+
+        List<String> args = spawnedArgs(herdr);
+        int bytes = "claude".length();
+        for (String arg : args) {
+            bytes += arg.getBytes(java.nio.charset.StandardCharsets.UTF_8).length + 3;
+        }
+        assertTrue(bytes <= HerdrPeerLauncher.PANE_COMMAND_BYTE_LIMIT,
+                "the launch command must fit the pane line: " + bytes + " bytes vs limit "
+                        + HerdrPeerLauncher.PANE_COMMAND_BYTE_LIMIT + " — args " + args);
+    }
+
+    /**
+     * fleetd #220: the guard refuses a command that cannot fit, instead of letting the pty drop the
+     * tail. The refusal must name the size and the argument to blame — a spawn that fails with
+     * "did not reach injectable state" tells the operator nothing, which is the whole reason this
+     * bug took a live pane scrape to find.
+     */
+    @Test
+    void anOverlongLaunchCommandIsRefusedWithTheSizeAndTheCulprit() {
+        FakeHerdr herdr = new FakeHerdr();
+        String huge = "x".repeat(1500);
+        ClaudeCodeLauncher launcher = service(herdr, List.of("claude", huge), null);
+
+        PeerUnreachableException refused = assertThrows(PeerUnreachableException.class, launcher::spawn);
+
+        assertTrue(refused.getMessage().contains("1024"), "names the limit: " + refused.getMessage());
+        assertTrue(refused.getMessage().contains("1500"), "names the culprit's size: " + refused.getMessage());
+        assertFalse(herdr.called("agent.start"),
+                "nothing may be started — a truncated command is worse than no spawn");
     }
 
     // CB-634: a profile with ideMcpUrl set mounts the IDE Index MCP as a second server and pins
@@ -262,10 +332,16 @@ class ClaudeCodeLauncherTest {
 
         Map<?, ?> start = (Map<?, ?>) herdr.lastCall("agent.start").params();
         assertEquals("claude", start.get("kind"), "herdr launches the canonical executable by kind");
-        // CB-533: the shared fixture pins model "coder", so the model flag is the whole args list.
         // What this test guards is that argv[0] is NOT repeated — herdr supplies it from `kind`.
-        assertEquals(List.of("--model", "coder"), start.get("args"),
-                "the configured executable is not repeated in args");
+        // fleetd #214: a plain spawn now also carries fleetd's own minted session id.
+        List<String> args = spawnedArgs(herdr);
+        assertFalse(args.contains("claude"), "the configured executable is not repeated in args");
+        int flag = args.indexOf("--session-id");
+        assertTrue(flag >= 0, "a plain spawn mints a session id: " + args);
+        assertDoesNotThrow(() -> UUID.fromString(args.get(flag + 1)), "the minted id is a valid UUID");
+        assertEquals(List.of("--model", "coder"),
+                List.of(args.get(args.size() - 2), args.get(args.size() - 1)),
+                "the CB-533 model flag still trails the launch flags: " + args);
     }
 
     @Test
@@ -278,7 +354,15 @@ class ClaudeCodeLauncherTest {
         assertFalse(args.contains("--append-system-prompt"), "no reply charter without mcpUrl");
         // CB-533: the model flag is independent of the MCP mount — pinning the model is not part of
         // "mount the bridge", so an unmounted worker still runs the model its profile names.
-        assertEquals(List.of("--verbose", "--model", "coder"), args,
+        // fleetd #214: a plain spawn now carries fleetd's own minted session id; drop the two
+        // mint elements when checking the rest of the argv.
+        int flag = args.indexOf("--session-id");
+        assertTrue(flag >= 0, "a plain spawn mints a session id even without a bridge mount: " + args);
+        assertDoesNotThrow(() -> UUID.fromString(args.get(flag + 1)), "the minted id is a valid UUID");
+        List<String> rest = new java.util.ArrayList<>(args);
+        rest.remove(flag + 1);
+        rest.remove(flag);
+        assertEquals(List.of("--verbose", "--model", "coder"), rest,
                 "the operator's own args are preserved, in order, ahead of the model flag");
     }
 
@@ -649,7 +733,7 @@ class ClaudeCodeLauncherTest {
         assertEquals("/work/proj", cwd, "effectiveCwd via SpawnRequest must match the three-arg resolution");
     }
 
-    // --- CB-547a: durable session identity (mint / resume / no-identity legacy) -----------------
+    // --- CB-547a / fleetd #214: durable session identity (always mint / resume) -----------------
 
     @Test
     void freshSpawnMintsASessionIdAndPassesTheName() {
@@ -684,18 +768,28 @@ class ClaudeCodeLauncherTest {
     }
 
     @Test
-    void noIdentitySpawnKeepsTheLegacyArgvAndCarriesNoSessionHandle() {
+    void plainSpawnMintsASessionIdSoEveryMemberIsResumable() {
+        // fleetd #214: a plain spawn passes no sessionName and no resumeSessionId, yet the member
+        // must still be resumable — the id is minted unconditionally, and it is the ONLY resume
+        // handle a claude-code member has (unlike opencode, nothing resolves it after the launch).
+        // The binary requires a valid UUID (checked against claude 2.1.252: a non-UUID is refused
+        // at argument parsing with "Invalid session ID. Must be a valid UUID.").
         FakeHerdr herdr = new FakeHerdr();
         ClaudeCodeLauncher svc = service(herdr, List.of("ccs", "ltms-local"), null);
 
         PeerHandle handle = svc.spawn(new SpawnRequest("ltms-local", null, null));
 
         List<String> args = spawnedArgs(herdr);
-        assertFalse(args.contains("--session-id"), "no identity → no --session-id");
-        assertFalse(args.contains("-n"), "no identity → no -n");
-        assertFalse(args.contains("-r"), "no identity → no -r");
-        assertNull(handle.agentSessionId(), "no identity → no resume handle");
-        assertNull(handle.sessionName(), "no identity → no logical name");
+        int flag = args.indexOf("--session-id");
+        assertTrue(flag >= 0 && flag + 1 < args.size(),
+                "--session-id is minted even when no identity is requested: " + args);
+        String minted = args.get(flag + 1);
+        assertDoesNotThrow(() -> UUID.fromString(minted), "--session-id is a valid UUID: " + minted);
+        assertEquals(minted, handle.agentSessionId(),
+                "the resume handle is the minted id, so every member is resumable from fleet_list");
+        assertFalse(args.contains("-n"), "no sessionName was requested → no -n");
+        assertFalse(args.contains("-r"), "no resume was requested → no -r");
+        assertNull(handle.sessionName(), "no sessionName was requested → no logical name");
     }
 
     @Test
@@ -869,6 +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
@@ -1660,4 +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

@@ -1,30 +1,43 @@
 package dev.ltms.fleet.member;
 
+import ch.qos.logback.classic.Level;
+import ch.qos.logback.classic.Logger;
+import ch.qos.logback.classic.spi.ILoggingEvent;
+import ch.qos.logback.core.read.ListAppender;
 import com.fasterxml.jackson.databind.JsonNode;
 import com.fasterxml.jackson.databind.ObjectMapper;
 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.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;
 
+import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
+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.function.Supplier;
 
 import static org.junit.jupiter.api.Assertions.*;
+import static org.junit.jupiter.api.Assumptions.assumeTrue;
 
 /**
  * The opencode adapter's launch build: a file-based MCP mount + reply-charter instructions (no
@@ -383,6 +396,46 @@ 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();
@@ -610,4 +663,474 @@ class OpenCodeLauncherTest {
         assertTrue(json.path("mcp").path("intellij").isMissingNode(),
                 "no IDE server when ideMcpUrl is unset");
     }
+
+    // --- fleetd #219: config root + discovery root under memberHerdrSocket ------------------------
+
+    /** 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 OpenCodeLauncher serviceWithConfig(FakeHerdr herdr, Path configRoot, Path discoveryRoot,
+                                                        FleetConfig.Profile cfg, Supplier<FleetConfig> config) {
+        return new OpenCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null,
+                0, System::currentTimeMillis, () -> { }, configRoot, discoveryRoot, 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 #219 site 1, acceptance criterion 1: with {@code memberHerdrSocket} configured and both
+     * {@code worktreeRoot}/{@code worktreeGroup} set, the generated {@code opencode.json} directory
+     * lives under {@code worktreeRoot} — NEVER under the injected {@code configRoot} (standing in for
+     * {@code java.io.tmpdir}, fleetd's own 0700 temp dir, unreadable by the member's different OS
+     * user) — and is shared read-only with the group via the SAME mechanism (fleetd #213's {@link
+     * EnvAllowListScrub#shareWithGroup}) the ZDOTDIR scrub uses.
+     */
+    @Test
+    void memberHerdrSocketWithWorktreeRootAndGroupPutsConfigDirUnderWorktreeRootAndSharesIt(
+            @TempDir Path configRoot, @TempDir Path worktreeRoot) throws Exception {
+        String group = currentUserGroup();
+        FakeHerdr herdr = new FakeHerdr();
+        serviceWithConfig(herdr, configRoot, configRoot,
+                opencodeCfg("google/gemini-2.5-pro", "http://127.0.0.1:8765/mcp", null),
+                () -> configWithMemberHerdrSocket(worktreeRoot.toString(), group)).spawn();
+
+        String cfgPath = startEnv(herdr).get("OPENCODE_CONFIG");
+        assertNotNull(cfgPath, "the profile still needs a config file");
+        Path cfgFile = Path.of(cfgPath);
+        Path dir = cfgFile.getParent();
+        assertEquals(worktreeRoot.toAbsolutePath().normalize(), dir.getParent(),
+                "the generated directory's parent must be worktreeRoot, not the injected configRoot "
+                        + "standing in for java.io.tmpdir — got parent " + dir.getParent());
+        assertFalse(dir.startsWith(configRoot),
+                "the generated directory must NOT be created under configRoot when memberHerdrSocket "
+                        + "is configured: " + dir);
+
+        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(cfgFile)),
+                "opencode.json must be group-readable, never group-writable");
+    }
+
+    /**
+     * fleetd #219 site 1, acceptance criterion 2: with {@code memberHerdrSocket} configured but
+     * NEITHER {@code worktreeRoot} nor {@code worktreeGroup} set, the launcher must refuse the spawn
+     * rather than write a config under {@code java.io.tmpdir} the member cannot read — that member
+     * would occupy a pane and never become deliverable, with nothing pointing at the real cause.
+     */
+    @Test
+    void memberHerdrSocketWithoutWorktreeRootOrGroupRefusesTheSpawn(@TempDir Path configRoot) {
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = serviceWithConfig(herdr, configRoot, configRoot,
+                opencodeCfg("google/gemini-2.5-pro", "http://127.0.0.1:8765/mcp", null),
+                () -> configWithMemberHerdrSocket(null, null));
+
+        IllegalStateException ex = assertThrows(IllegalStateException.class,
+                () -> launcher.spawn(new SpawnRequest(null, null, null)),
+                "a missing worktreeRoot/worktreeGroup must refuse the spawn, not write an unreadable config");
+        assertTrue(ex.getMessage().contains("worktreeRoot"),
+                "the refusal must name the missing config key — got: " + ex.getMessage());
+        assertFalse(herdr.called("tab.create"),
+                "the spawn must be refused BEFORE any pane is created — got calls: " + herdr.calls);
+    }
+
+    /**
+     * fleetd #219 site 1, 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 directory.
+     */
+    @Test
+    void memberHerdrSocketWithWorktreeRootButNoGroupRefusesTheSpawn(
+            @TempDir Path configRoot, @TempDir Path worktreeRoot) {
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = serviceWithConfig(herdr, configRoot, configRoot,
+                opencodeCfg("google/gemini-2.5-pro", "http://127.0.0.1:8765/mcp", null),
+                () -> configWithMemberHerdrSocket(worktreeRoot.toString(), null));
+
+        IllegalStateException ex = assertThrows(IllegalStateException.class,
+                () -> launcher.spawn(new SpawnRequest(null, null, null)));
+        assertTrue(ex.getMessage().contains("worktreeGroup"),
+                "worktreeRoot alone is not enough — got: " + ex.getMessage());
+    }
+
+    /**
+     * fleetd #219 site 1, 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 generated directory must still
+     * land directly under the injected {@code configRoot}, byte-identical to before this fix.
+     */
+    @Test
+    void memberHerdrSocketAbsentStaysUnderConfigRootEvenWithALiveConfigSupplier(@TempDir Path configRoot)
+            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, configRoot, configRoot,
+                opencodeCfg("google/gemini-2.5-pro", "http://127.0.0.1:8765/mcp", null), () -> config)
+                .spawn();
+
+        String cfgPath = startEnv(herdr).get("OPENCODE_CONFIG");
+        assertNotNull(cfgPath);
+        assertTrue(Path.of(cfgPath).startsWith(configRoot),
+                "with memberHerdrSocket absent, the config directory must still be created directly "
+                        + "under configRoot, unchanged from before this fix");
+    }
+
+    /**
+     * fleetd #219 site 2: under {@code memberHerdrSocket}, opencode session discovery must be
+     * declared unavailable rather than silently scanning fleetd's own {@code discoveryRoot} — which,
+     * under this config key, is NOT where the member's opencode actually writes its session
+     * database. This test proves the gate is real, not merely "no record yet": a matching record IS
+     * written to {@code discoveryRoot} (the exact fixture {@link
+     * #theHandleDiscoversTheSessionIdForTheWorkersCwdOnlyAfterItAppears} proves discovery would
+     * otherwise find), and {@code agentSessionId()} must still return {@code null} — proving the
+     * gate, not a coincidental absence of data, is what produced the null. One WARN is also logged,
+     * exactly once even across repeated calls.
+     */
+    @Test
+    void discoveryIsUnavailableUnderMemberHerdrSocketEvenWhenARecordExists(
+            @TempDir Path configRoot, @TempDir Path worktreeRoot, @TempDir Path discRoot) throws Exception {
+        String group = currentUserGroup();
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_should_be_hidden", "/work/dir", 1000L);
+
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = new OpenCodeLauncher(new AgentControl(herdr),
+                new WorkspaceControl(herdr), Map.of("gemini", opencodeCfg(null, null, null)),
+                "gemini", _ -> null, 0, System::currentTimeMillis, () -> { }, configRoot, discRoot,
+                null, null, () -> configWithMemberHerdrSocket(worktreeRoot.toString(), group));
+
+        Logger logger = (Logger) LoggerFactory.getLogger(OpenCodeLauncher.class);
+        Level original = logger.getLevel();
+        logger.setLevel(Level.INFO);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        PeerHandle handle;
+        try {
+            handle = launcher.spawn(new SpawnRequest(null, "/work/dir", null));
+            assertNull(handle.agentSessionId(),
+                    "memberHerdrSocket configured: discovery must stay unavailable even though a "
+                            + "matching record exists in discoveryRoot");
+            assertNull(handle.agentSessionId(), "the gate must hold on a second call too");
+        } finally {
+            logger.detachAppender(appender);
+            logger.setLevel(original);
+        }
+        List<String> warnings = appender.list.stream()
+                .filter(e -> e.getLevel() == Level.WARN)
+                .map(ILoggingEvent::getFormattedMessage)
+                .toList();
+        assertEquals(1, warnings.size(),
+                "exactly one WARN across two agentSessionId() calls — got: " + warnings);
+        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) -> 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) -> 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) -> 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) -> 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) -> 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) -> 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) -> 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) -> 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) -> 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/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 +145,96 @@ class OpenCodeSessionDiscoveryTest {
         assertNull(new OpenCodeSessionDiscovery(root).sessionIdForDirectory("/w/a"),
                 "an unreadable database resolves to null, not an exception");
     }
+
+    // --- fleetd #175: actualModelForDirectory / 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).actualModelForDirectory("/w/a");
+
+        assertNotNull(actual, "a well-formed model JSON parses");
+        assertEquals("gpt-5.6-terra", actual.id());
+        assertEquals("openai", actual.provider());
+    }
+
+    @Test
+    void prefersTheModelOfTheMostRecentlyUpdatedRow(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_old", "/w/a", 1000L, "{\"id\":\"old-model\",\"providerID\":\"openai\"}");
+        writeRecord(root, "ses_new", "/w/a", 5000L, "{\"id\":\"new-model\",\"providerID\":\"openai\"}");
+
+        assertEquals("new-model",
+                new OpenCodeSessionDiscovery(root).actualModelForDirectory("/w/a").id(),
+                "the model of the row with the highest time_updated wins, same as the id");
+    }
+
+    @Test
+    void aNonMatchingDirectoryYieldsUnknownModelRatherThanAMismatch(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, "{\"id\":\"x\",\"providerID\":\"y\"}");
+
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForDirectory("/w/other"),
+                "no row for this cwd yet → unknown, not a wrong model");
+    }
+
+    @Test
+    void aNullModelColumnYieldsUnknownWithoutThrowing(@TempDir Path root) throws Exception {
+        writeRecord(root, "ses_aaa", "/w/a", 1000L, null);
+
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForDirectory("/w/a"),
+                "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).actualModelForDirectory("/w/a"),
+                "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).actualModelForDirectory("/w/a"),
+                "no id in the JSON → unknown, since id is what a caller actually compares");
+    }
+
+    @Test
+    void aMissingDatabaseYieldsUnknownModelWithoutThrowing(@TempDir Path root) {
+        assertNull(new OpenCodeSessionDiscovery(root).actualModelForDirectory("/w/a"));
+    }
+
+    /**
+     * 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#actualModelForDirectory} 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.actualModelForDirectory("/w/a"),
+                "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
@@ -887,15 +1019,18 @@ class SessionManagerTest {
     }
 
     @Test
-    void acquireWithNeitherSessionFieldLeavesAgentSessionIdNull() {
+    void acquireWithNeitherSessionFieldStillMintsAnAgentSessionId() {
+        // fleetd #214: the claude-code launcher mints a session id for EVERY spawn, so a member is
+        // resumable even when the spawn asked for no session identity.
         FakeHerdr herdr = new FakeHerdr();
         SessionManager sessions = sessionManager(herdr);
 
         MemberSession s = sessions.acquire("ltms-local", null, null, null);
 
-        assertNull(s.agentSessionId(), "no identity requested — unchanged from before CB-584");
-        assertFalse(SessionManager.rosterView(s, null).containsKey("agentSessionId"),
-                "a null id is omitted from the roster, like charterSha256 for a receipt-less session");
+        assertNotNull(s.agentSessionId(),
+                "fleetd #214: a plain spawn mints a session id, so every member is resumable");
+        assertTrue(SessionManager.rosterView(s, null).containsKey("agentSessionId"),
+                "the minted id is in the roster, so fleet_list advertises every member's resume handle");
     }
 
     @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