fleetd #267: warn once per profile when the model check can't run

OpenCodeLauncher.SessionAwareHandle.agentSessionId() is the only caller of
checkModelMatch (fleetd #175), and it sits behind the fleetd #249 worktree
gate. A spawn with no worktree:true — the ordinary shape of most opencode
spawns — never reached the check at all, and the gap was totally silent.

The check cannot be decoupled from agentSessionId()'s resolved id: doing so
would re-derive 'whatever is newest in the shared directory' and reintroduce
the false-positive risk fleetd #234 fixed (a sibling's differently-configured
model looking like a mismatch for a profile that never actually ran it). The
#249 gate is correct and stays as-is.

Instead, log once per profile at WARN, naming the profile, the same
treatment discoveryUnavailable already gets a few lines above — a logged
UNKNOWN beats a check that silently never runs.

deploy/fleetd.service

@@ -31,11 +31,27 @@ Environment=HERDR_SOCKET_PATH=%h/.config/herdr/herdr.sock
 # spawns, so this line decides whether the fleet can run a build at all. systemd does not source a
 # login shell, so without it the daemon — and every worker — gets a bare default with no JDK/Maven.
 Environment=PATH=/usr/lib/jvm/temurin-25-jdk/bin:/usr/share/maven/bin:/usr/local/bin:/usr/bin:/bin
-# Secrets are NOT set here — this file is committed. Put the API/worker tokens in a private
-# drop-in that systemd reads with restrictive permissions:
-#   systemctl --user edit fleetd     →  [Service] / Environment=FLEETD_API_TOKEN=...
-# or point EnvironmentFile at a 0600 file:
+# Secrets are NOT set here — this file is committed. Put ALL three tokens in a private drop-in
+# that systemd reads with restrictive permissions. In `systemctl --user edit fleetd`, add:
+#   [Service]
+#   Environment=FLEETD_API_TOKEN=...
+#   Environment=WORKER_GITEA_TOKEN=...
+#   Environment=AI_GATEWAY_TOKEN=...
+# FLEETD_API_TOKEN protects fleetd's API. WORKER_GITEA_TOKEN lets members open pull requests; if
+# it is missing, fleetd still starts, but a member fails when it later tries to open a pull request.
+# AI_GATEWAY_TOKEN authenticates gateway profiles; if it is missing, fleetd still starts, but a
+# gateway profile later returns HTTP 401. Or, put the same three variables in a 0600 file and add:
 # EnvironmentFile=%h/.config/fleetd/env
+# After starting, check which of them actually resolved. The daemon reports every secret a
+# configured profile references, by name, never by value:
+#   journalctl --user -u fleetd | grep 'startup secret'
+# A resolved one logs "startup secret NAME: set (profile 'x' tokenEnv)". A missing one logs
+# "startup secret NAME: MISSING" at WARN — and the daemon starts anyway, which is the whole
+# problem: without this grep the first sign is a member that cannot open a pull request, hours
+# later and in a different component.
+# Note what the report can and cannot tell you. It lists only names some profile actually
+# references (tokenEnv, gitTokenEnv, and the broker uriEnv). A secret nothing references is never
+# reported, because nothing needs it.
 
 Restart=on-failure
 RestartSec=10s

fleetd/fleetd.example.yaml

@@ -307,16 +307,22 @@ profiles:
     #   and no refusal on cost; the cap on live members is the single thing standing between a
     #   fan-out and your monthly limit. Set it deliberately and keep it small.
     #
-    #   GOTCHA 3 (fleetd #176) — `maxLoad` counts members, never the lead itself. The lead is a live
-    #   `claude` session on this SAME account (a lead is never moved off-subscription, whatever its
-    #   own profile says), so it already holds one seat before any member spawns. If a lead's
-    #   `fleet.leaders.<name>.profile` names THIS profile — or ANY OTHER `subscription: true`
-    #   profile that shares this one's account (see THE SENTINEL, just below, next to
-    #   `credentialId:`) — `fleet_list`'s `free` for this profile subtracts that lead's live
-    #   seat(s) automatically; see `profile:` under THE FLEET below. If no lead entry names a
-    #   profile sharing this account, fleetd has no way to know a lead holds a seat here, and `free`
-    #   will overstate what a fresh `fleet_spawn` actually gets by exactly the seats the lead is
-    #   quietly holding.
+    #   GOTCHA 3 (fleetd #176, corrected by fleetd #257) — `maxLoad` counts members, never the lead
+    #   itself. The lead is a live `claude` session on this SAME account (a lead is never moved
+    #   off-subscription, whatever its own profile says), so it already holds one seat before any
+    #   member spawns. If a lead's `fleet.leaders.<name>.profile` names THIS profile — or ANY OTHER
+    #   `subscription: true` profile that shares this one's account (see THE SENTINEL, just below,
+    #   next to `credentialId:`) — `fleet_list` reports that seat count under `leadSeats`; see
+    #   `profile:` under THE FLEET below. `free` itself is NEVER reduced by `leadSeats`: `free` means
+    #   "what the real placement gate (`CompositePeerLauncher#enforceMaxLoad`) will actually grant a
+    #   fresh `fleet_spawn` right now", and that gate only ever compares live members against
+    #   `maxLoad` — it has no notion of the lead's own seat. An earlier cut of this feature
+    #   subtracted `leadSeats` from `free` on the theory it made `free` describe the true ceiling on
+    #   the account, but no backend seat ceiling shared with the lead has ever actually been
+    #   measured, and the subtraction just made `free` disagree with the one thing it is supposed to
+    #   describe — the fleetd #257 fix. `maxLoad: 3` means 3 member slots, full stop; a lead sharing
+    #   the account is a fact you can see in `leadSeats`, not a reason `free` undercounts spawns that
+    #   will, in practice, succeed.
     #
     #   THE SENTINEL (fleetd #176 stage 2, correcting an inert stage 1 fix): every `subscription:
     #   true` profile that leaves `credentialId` unset shares ONE implicit account-wide credential
@@ -536,17 +542,18 @@ fleet:
   # auto-launched: it is also how fleetd learns which account this lead's own session shares. A
   # `subscription: true` profile bills the operator's Claude account, and the lead itself is always
   # a live `claude` session on that same account — `maxLoad` never counted that seat. If a lead
-  # entry here names a profile that shares a worker profile's account, `fleet_list`'s `free` for
-  # that worker profile subtracts the lead's live seat(s) automatically. "Shares the account" is
-  # decided by matching `effectiveCredentialId()`, which (fleetd #176 stage 2 — see THE SENTINEL,
-  # next to `credentialId:`, in THE WORKERS above) means: an explicit, matching `credentialId:` on
-  # both, OR — the common case, needing NO extra config — both being `subscription: true` with
-  # `credentialId` left unset, since those all share one implicit account-wide id. A lead on `opus`
-  # and workers on `sonnet` link automatically this way; they do NOT need the same profile name.
-  # Setting `profile:` on an already-running, recognise-only lead is safe — the daemon only launches
-  # the SHORTFALL below `instances`, so naming a profile here does not, by itself, start anything.
-  # Omit it and fleetd has no way to derive the sharing — there is no other reliable signal on the
-  # daemon's side — so that lead's seat goes uncounted, exactly as before this ticket.
+  # entry here names a profile that shares a worker profile's account, `fleet_list` reports the
+  # lead's live seat(s) on that worker profile under `leadSeats` — informational only, as of fleetd
+  # #257 it is NEVER subtracted from `free` (see GOTCHA 3, next to `maxLoad:`, in THE WORKERS above,
+  # for why). "Shares the account" is decided by matching `effectiveCredentialId()`, which (fleetd
+  # #176 stage 2 — see THE SENTINEL, next to `credentialId:`, in THE WORKERS above) means: an
+  # explicit, matching `credentialId:` on both, OR — the common case, needing NO extra config — both
+  # being `subscription: true` with `credentialId` left unset, since those all share one implicit
+  # account-wide id. A lead on `opus` and workers on `sonnet` link automatically this way; they do
+  # NOT need the same profile name. Setting `profile:` on an already-running, recognise-only lead is
+  # safe — the daemon only launches the SHORTFALL below `instances`, so naming a profile here does
+  # not, by itself, start anything. Omit it and fleetd has no way to derive the sharing — there is
+  # no other reliable signal on the daemon's side — so that lead's seat never appears in `leadSeats`.
   #
   # `tab:` (CB-579) is REQUIRED and is the only field identity depends on — the exact label of the
   # tab hosting the lead, matched case-insensitively. Label the tab yourself and put that same
@@ -669,20 +676,35 @@ guard:
 #            every name here NOT also in `allow` is overlaid with a non-secret sentinel value before
 #            the pane's login shell runs — real protection only for names that shell does not itself
 #            re-export (see the ROUND-2 CORRECTION note above). Under allow-list: reporting only.
-#   sshAuthSock → whether SSH_AUTH_SOCK may pass through under allow-list ("allow") or must be
-#            blanked like any other non-derived name ("block", the default). This is a decision you
-#            have to make explicitly: SSH_AUTH_SOCK is a handle to YOUR ssh-agent, and a member
-#            holding it can sign with your keys — it sits in no secret file and looks like no
-#            credential, which is why it slipped past three earlier tickets (gitea #110). Blocking
-#            it breaks git over SSH inside members (push/fetch authenticate as you); use HTTPS
-#            remotes or scoped deploy keys instead of allowing it lightly.
+#   sshAgentEnv → whether SSH_AUTH_SOCK may pass through under allow-list ("inherit") or is omitted
+#            from the member environment ("omit", the default). Omitting it only omits the
+#            inherited ssh-agent path. It discourages automatic use of the operator's agent.
+#            It does not deny same-user access to that socket. It also does not block SSH keys that
+#            are readable on disk. Git over SSH may still work from inside a member. Keep the block:
+#            it is correct and costs nothing, but it is not a control. A member runs as the same OS
+#            user as the lead. Inside one uid, ordinary Unix permissions provide no meaningful
+#            confidentiality boundary. A real boundary needs a different OS user or OS-level
+#            confinement, such as a container or VM. That is the open question in fleetd #184.
+#
+#            Still do not set this to "inherit" casually. SSH_AUTH_SOCK is a live handle to YOUR
+#            ssh-agent, so a member holding it can sign with EVERY key the agent holds. It sits in
+#            no secret file and looks like no credential, which is why it slipped past three
+#            earlier tickets (gitea #110). Blocking it does not contain a member, but allowing it
+#            hands one a signing capability for no gain — the block costs nothing, so keep it.
+#
+#            Both halves of this are measured, not argued. 2026-08-28: a member with
+#            SSH_AUTH_SOCK blanked pushed to the forge over SSH successfully, because `ssh -G`
+#            resolves an IdentityFile outside ~/.ssh that is readable and has no passphrase. An
+#            earlier version of this comment claimed blocking the socket BREAKS git over SSH. It
+#            does not. That claim came from looking only in ~/.ssh, which holds nothing but four
+#            `Include` lines — looking in one place and concluding about the whole host.
 #
 # HOT-RELOADABLE the same way `fleet:` is (CB-559): read fresh on every spawn, so editing this list
 # and reloading config (or restarting) changes what the NEXT spawn inherits; already-running members
 # are unaffected either way.
 # memberCredentials:
 #   policy: deny-by-default        # or "deny-list", or "allow-list" (CB-633) — see above
-#   sshAuthSock: block             # allow-list only; see the sshAuthSock note above
+#   sshAgentEnv: omit              # allow-list only; see the sshAgentEnv note above
 #   allow:
 #     - AI_GATEWAY_TOKEN     # named in a profile's tokenEnv (local/gx) — a member reaching the
 #                            # gateway is by design, not a leak

fleetd/pom.xml

@@ -29,6 +29,7 @@
         <commons-compress.version>1.27.1</commons-compress.version>
         <commons-lang3.version>3.18.0</commons-lang3.version>
         <sqlite-jdbc.version>3.53.4.0</sqlite-jdbc.version>
+        <archunit.version>1.5.0</archunit.version>
     </properties>
 
     <!--
@@ -176,6 +177,14 @@
             <version>${testcontainers.version}</version>
             <scope>test</scope>
         </dependency>
+
+        <!-- fleetd #131: package-boundary and cycle enforcement (PackageCyclesTest). -->
+        <dependency>
+            <groupId>com.tngtech.archunit</groupId>
+            <artifactId>archunit-junit5</artifactId>
+            <version>${archunit.version}</version>
+            <scope>test</scope>
+        </dependency>
     </dependencies>
 
     <build>

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

@@ -121,6 +121,7 @@ public final class Fleetd {
         // else can fail on a silently-empty one. A daemon started without a login shell (launchd)
         // boots fine either way — this is the only thing that says so out loud.
         reportRequiredSecrets(cfg);
+        reportMemberTrustModel(cfg);
         // CB-596: an absent (or empty) memberCredentials: block blocks NOTHING — no credential
         // name is hardcoded any more to fall back on. Say so loudly, the same way a missing
         // secret is reported above, so upgrading past this commit never silently drops CB-592's
@@ -1136,6 +1137,24 @@ public final class Fleetd {
         });
     }
 
+    /**
+     * fleetd #184: state the member trust model at startup. Environment controls and worktrees do
+     * not make a sandbox when fleetd and its members use the same OS user. A separate herdr may
+     * provide that boundary, but fleetd cannot inspect the uid at the other end of its socket.
+     */
+    static void reportMemberTrustModel(FleetConfig cfg) {
+        if (cfg.memberHerdrSocket() != null && !cfg.memberHerdrSocket().isBlank()) {
+            log.info("member trust model: members are routed to a separate herdr through "
+                    + "memberHerdrSocket. fleetd cannot see that herdr's uid, so confirm it runs "
+                    + "as a different OS user before treating it as a boundary.");
+            return;
+        }
+        log.info("member trust model: members run as the same OS user as fleetd, not in a sandbox. "
+                + "A member can read any file this user can read, including SSH keys and credential "
+                + "stores, whatever memberCredentials says. To add a real boundary, route members to "
+                + "a second herdr under a different OS user with memberHerdrSocket.");
+    }
+
     /**
      * CB-596: {@code known:} empty (block absent entirely, or present but empty) means {@link
      * FleetConfig.MemberCredentials#blockedSet()} is empty too — every member pane inherits the

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

@@ -1,6 +1,8 @@
 package dev.ltms.fleet.config;
 
+import com.fasterxml.jackson.annotation.JsonCreator;
 import com.fasterxml.jackson.annotation.JsonIgnoreProperties;
+import com.fasterxml.jackson.annotation.JsonProperty;
 import com.fasterxml.jackson.core.JsonParser;
 import com.fasterxml.jackson.core.JsonToken;
 import com.fasterxml.jackson.databind.ObjectMapper;
@@ -1357,19 +1359,19 @@ public record FleetConfig(
      *               deny-list/deny-by-default every name here that is NOT also in {@link #allow} is
      *               overlaid with a non-secret sentinel value. Under allow-list this list is
      *               reporting only.
-     * @param sshAuthSock whether the member may inherit {@code SSH_AUTH_SOCK} under the allow-list
-     *               policy ({@code "allow"}) or must have it blanked ({@code "block"}, the default).
+     * @param sshAgentEnv whether the member may inherit {@code SSH_AUTH_SOCK} under the allow-list
+     *               policy ({@code "inherit"}) or must have it omitted ({@code "omit"}, the default).
      *               This is a DECISION, never a default: {@code SSH_AUTH_SOCK} is a handle to the
      *               operator's ssh-agent, and a member holding it can sign with the operator's own
      *               keys — but it appears in no secret file and is credential-shaped like nothing on
      *               any list, which is why three earlier tickets missed it (gitea #110 / CB-607).
-     *               Blocking it breaks git over SSH inside the member; allow it only when members do
+     *               Omitting it does not prevent git over SSH inside the member; inherit it only when members do
      *               not need to authenticate as the operator over SSH. Ignored under deny-list /
      *               deny-by-default, which never touch the name.
      */
     @JsonIgnoreProperties(ignoreUnknown = true)
     public record MemberCredentials(String policy, List<String> allow, List<String> known,
-                                    String sshAuthSock) {
+                                     String sshAgentEnv) {
 
         /** Default policy: block every {@code known} name not in {@code allow}, via the env overlay. */
         public static final String POLICY_DENY_BY_DEFAULT = "deny-by-default";
@@ -1386,11 +1388,25 @@ public record FleetConfig(
          */
         public static final String POLICY_ALLOW_LIST = "allow-list";
 
-        /** The pre-CB-633 three-field form — {@code sshAuthSock} defaults to blocked. */
+        /** The pre-CB-633 three-field form — {@code sshAgentEnv} defaults to omitted. */
         public MemberCredentials(String policy, List<String> allow, List<String> known) {
             this(policy, allow, known, null);
         }
 
+        /**
+         * Reads both the current {@code sshAgentEnv} key and the compatible {@code sshAuthSock} key.
+         * When both keys are present, {@code sshAgentEnv} wins, even if its value is unrecognised.
+         */
+        @JsonCreator
+        public static MemberCredentials fromYaml(@JsonProperty("policy") String policy,
+                                                @JsonProperty("allow") List<String> allow,
+                                                @JsonProperty("known") List<String> known,
+                                                @JsonProperty("sshAgentEnv") String sshAgentEnv,
+                                                @JsonProperty("sshAuthSock") String sshAuthSock) {
+            return new MemberCredentials(policy, allow, known,
+                    sshAgentEnv != null ? sshAgentEnv : sshAuthSock);
+        }
+
         public MemberCredentials {
             String normalizedPolicy = (policy == null || policy.isBlank())
                     ? POLICY_DENY_BY_DEFAULT : policy.toLowerCase(java.util.Locale.ROOT);
@@ -1399,8 +1415,9 @@ public record FleetConfig(
             policy = POLICY_DENY_LIST.equals(normalizedPolicy) ? POLICY_DENY_BY_DEFAULT : normalizedPolicy;
             allow = allow == null ? List.of() : List.copyOf(allow);
             known = known == null ? List.of() : List.copyOf(known);
-            sshAuthSock = (sshAuthSock != null && "allow".equalsIgnoreCase(sshAuthSock.trim()))
-                    ? "allow" : "block";
+            sshAgentEnv = (sshAgentEnv != null
+                    && ("inherit".equalsIgnoreCase(sshAgentEnv.trim())
+                    || "allow".equalsIgnoreCase(sshAgentEnv.trim()))) ? "inherit" : "omit";
         }
 
         /** True when this block selects the CB-633 derived-allow-list policy. */
@@ -1409,8 +1426,8 @@ public record FleetConfig(
         }
 
         /** True when {@code SSH_AUTH_SOCK} may pass through under the allow-list policy. Default: no. */
-        public boolean sshAuthSockAllowed() {
-            return "allow".equals(sshAuthSock);
+        public boolean sshAgentEnvInherited() {
+            return "inherit".equals(sshAgentEnv);
         }
 
         /** {@link #allow} as a set, for membership checks. */

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

@@ -139,16 +139,24 @@ public final class FleetMcp {
 
     /**
      * fleetd #176: the seats a profile's own live LEAD session(s) hold on the same Claude
-     * subscription — the third reason (alongside {@link QuarantineSource} and {@link OutageSource})
-     * {@code free} can overstate what a fresh {@code fleet_spawn} would actually get.
+     * subscription — a fact {@code fleet_list} reports alongside {@code free} via the
+     * {@code leadSeats} key.
      *
      * <p>{@code maxLoad} counts only <em>members</em>, never the lead itself. But a
      * {@code subscription: true} profile bills the operator's own Claude account, and the lead is
      * always a live {@code claude} session on that same account (it is never moved off-subscription
-     * — see {@code LeadLauncher}). So a fan-out that fills every member slot still leaves the lead's
-     * own seat unaccounted for, and the daemon reports a slot that was never really free. See
-     * {@code Fleetd.leadSeatLookup} for how the count is derived — from {@code fleet.leaders.<name>
-     * .profile} and each profile's {@code effectiveCredentialId()}, never a hardcoded constant.
+     * — see {@code LeadLauncher}). See {@code Fleetd.leadSeatLookup} for how the count is derived —
+     * from {@code fleet.leaders.<name>.profile} and each profile's {@code effectiveCredentialId()},
+     * never a hardcoded constant.
+     *
+     * <p>fleetd #257: this count is reported, never subtracted from {@code free}. An earlier cut of
+     * this feature subtracted it, on the theory that it made {@code free} describe the real ceiling
+     * on the account — but the real spawn gate ({@code CompositePeerLauncher#enforceMaxLoad}) never
+     * read this count at all, so the subtraction made {@code free} disagree with the one thing it is
+     * supposed to describe: what a fresh {@code fleet_spawn} will actually get. No backend seat
+     * ceiling shared with the lead has been measured either — see {@code fleetd.example.yaml}'s
+     * {@code maxLoad} docs. {@code free} now always equals {@code max(0, maxLoad - live)}, and
+     * {@code leadSeats} is reported purely as a fact the caller may act on however it likes.
      */
     public record LeadSeatSource(Function<String, Integer> seatsFor) {
         /** Inert source — no profile is ever reported as sharing a seat with a lead. */
@@ -1137,10 +1145,17 @@ public final class FleetMcp {
      * <p>fleetd #176: {@code maxLoad} counts panes, not subscription seats — it never counted the
      * lead's own seat on a {@code subscription: true} profile's account. {@link LeadSeatSource}
      * reports that count (0 for a non-subscription profile, or when no live lead shares its
-     * credential), and it is subtracted from {@code free} the same way {@code live} already is —
-     * {@code maxLoad} itself is left untouched, so the row still reports the configured cap. The
-     * {@code leadSeats} key is added only when the count is positive, for the same
-     * byte-identical-when-unused reason as the quarantine/cool-off keys above.
+     * credential) via the {@code leadSeats} key, added only when the count is positive, for the
+     * same byte-identical-when-unused reason as the quarantine/cool-off keys above.
+     *
+     * <p>fleetd #257: {@code leadSeatCount} is reported, never subtracted from {@code free}.
+     * {@code free} means "what a fresh {@code fleet_spawn} on this profile will actually get", and
+     * the real gate ({@code CompositePeerLauncher#enforceMaxLoad}) only ever compares {@code live}
+     * against {@code maxLoad} — it has no notion of a lead's own seat. Subtracting
+     * {@code leadSeatCount} here made {@code free} disagree with the gate it is supposed to
+     * describe: it could report {@code free: 0} while a spawn on that exact profile still
+     * succeeded. {@code leadSeats} stays in the row as a fact the caller can act on however it
+     * likes, but it no longer changes what {@code free} means.
      */
     private static Map<String, Object> capacityView(String profile, Function<String, Integer> liveCount,
                                                      Function<String, Integer> maxLoad, List<MemberSession> roster,
@@ -1155,7 +1170,12 @@ public final class FleetMcp {
                 .count();
         Map<String, Object> row = new LinkedHashMap<>();
         row.put("profile", profile); row.put("maxLoad", cap); row.put("live", live);
-        row.put("free", cap == null ? null : Math.max(0, cap - live - leadSeatCount));
+        // fleetd #257: free must report what the real spawn gate (CompositePeerLauncher#enforceMaxLoad)
+        // will actually grant, and that gate never reads leadSeatCount — only maxLoad and live. Not
+        // subtracting the lead's seat here used to make free UNDERSTATE what a fresh fleet_spawn would
+        // get, so a lead believing free:0 gave up on a profile the gate would still spawn onto.
+        // leadSeatCount is still reported via the leadSeats key below, just never subtracted from free.
+        row.put("free", cap == null ? null : Math.max(0, cap - live));
         row.put("reclaimable", reclaimable);
         if (leadSeatCount > 0) {
             row.put("leadSeats", leadSeatCount);
@@ -1360,8 +1380,13 @@ public final class FleetMcp {
                         + "member spawned without one shares its directory with others and never reports "
                         + "an id, however long it runs (fleetd #249). An empty 'members' "
                         + "means no members are spawned; it says nothing about peers. When capacity "
-                        + "facts are configured, a 'capacity' row per profile also reports free: 0 for "
-                        + "a quarantined profile's credential (see fleet_profiles), whatever its "
+                        + "facts are configured, a 'capacity' row per profile reports 'free' — the "
+                        + "slots a fresh fleet_spawn on that profile will actually be granted right "
+                        + "now (max(0, maxLoad - live)), the same check the spawn gate itself runs. A "
+                        + "'leadSeats' key, when present, reports how many of those live slots are a "
+                        + "lead session sharing this profile's subscription — informational only, "
+                        + "already NOT subtracted from 'free' (fleetd #257). It also reports free: 0 "
+                        + "for a quarantined profile's credential (see fleet_profiles), whatever its "
                         + "maxLoad/live — with credentialId and quarantinedForSeconds naming the "
                         + "quarantine, so 'free: 0, busy' can be told apart from 'free: 0, refusing "
                         + "for N seconds'.",

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

@@ -19,6 +19,7 @@ import java.nio.file.Files;
 import java.nio.file.Path;
 import java.nio.file.StandardCopyOption;
 import java.nio.file.attribute.PosixFileAttributeView;
+import java.util.Arrays;
 import java.util.EnumSet;
 import java.util.List;
 import java.util.Map;
@@ -519,6 +520,26 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
      * the first. Both exist because of a real incident: see {@link HerdrPeerLauncher#isProvisionedWorktree}'s javadoc
      * and {@link #writeAtomically}'s javadoc.
      *
+     * <p><b>Compare-and-swap against a writer the lock cannot reach (fleetd #247).</b>
+     * {@code TRUST_JSON_LOCK} only serialises calls this launcher itself makes inside this one JVM.
+     * It does nothing about a writer outside it — and on a host where the profile's
+     * {@code configDir} is the operator's own {@code CLAUDE_CONFIG_DIR}, the file this method writes
+     * IS the operator's own live Claude Code session's config file, being read and written by that
+     * session while it runs. Measured 2026-09-03: its mtime moved minutes after a spawn while that
+     * session was active. A plain read-modify-write there is a routine lost update, not a rare one:
+     * fleetd reads v1, the operator's session reads v1 and writes v2 with their own change, fleetd's
+     * {@code ATOMIC_MOVE} then lands v3 built from v1 — atomic, but v2's change is gone. So before
+     * the move this method re-reads {@code target}'s exact bytes and compares them with the bytes it
+     * built its update from; a mismatch means someone else wrote in between, and it discards its
+     * work and rebuilds from the fresh bytes, up to {@link #MAX_TRUST_JSON_CAS_ATTEMPTS} times.
+     * <b>Exhausting the retries writes nothing</b> — see the WARN at the end of the loop for why
+     * that, not a last write-anyway, is the safe failure: the member shows the trust dialog and
+     * fails to reach an injectable state, which is visible, logged and recoverable; overwriting the
+     * operator's live config with a stale copy is neither. This narrows the lost-update window, it
+     * does not close it — a write landing between the final re-read and the {@code ATOMIC_MOVE}
+     * itself is still lost, because there is no OS-level compare-and-swap on a plain file, only this
+     * cooperative narrowing of the gap.
+     *
      * @param configDir the profile's {@code CLAUDE_CONFIG_DIR} ({@code cfg.configDir()}), or
      *                  {@code null}/blank to target the default {@code ~/.claude.json}
      * @param cwd       the spawn's resolved working directory — the exact key Claude Code will look
@@ -528,55 +549,118 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
         if (!isProvisionedWorktree(cwd)) {
             return;
         }
-        Path target = (configDir == null || configDir.isBlank())
+        boolean unsetConfigDir = configDir == null || configDir.isBlank();
+        Path target = unsetConfigDir
                 ? Path.of(System.getProperty("user.home"), ".claude.json")
                 : Path.of(configDir, ".claude.json");
+        if (unsetConfigDir) {
+            // fleetd #247: configDir unset is the ONLY path that targets ~/.claude.json — the
+            // operator's own home file, not a per-profile one — and it is the default, so a
+            // profile that simply forgot to set configDir gets no signal at all short of the
+            // operator noticing their own file changing. Say so loudly, every time it is about to
+            // happen, rather than only once ever: each occurrence is a live write to a real
+            // person's home config and deserves its own log line.
+            log.warn("seedTrustDialog: profile has no configDir set, so the workspace-trust seed "
+                    + "for cwd '{}' is about to write the operator's own default '{}' — set "
+                    + "configDir on this profile to target a per-member config file instead",
+                    cwd, target);
+        }
         synchronized (TRUST_JSON_LOCK) {
             try {
                 if (target.getParent() != null) {
                     Files.createDirectories(target.getParent());
                 }
-                ObjectNode root = null;
-                if (Files.isRegularFile(target)) {
-                    JsonNode existing = TRUST_JSON.readTree(target.toFile());
-                    if (existing instanceof ObjectNode existingObject) {
-                        root = existingObject;
+                for (int attempt = 1; attempt <= MAX_TRUST_JSON_CAS_ATTEMPTS; attempt++) {
+                    byte[] before = Files.isRegularFile(target) ? Files.readAllBytes(target) : null;
+                    ObjectNode root = parseTrustJsonOrEmpty(before);
+                    JsonNode projectsNode = root.get("projects");
+                    ObjectNode projects = projectsNode instanceof ObjectNode projectsObject
+                            ? projectsObject : TRUST_JSON.createObjectNode();
+                    if (!(projectsNode instanceof ObjectNode)) {
+                        root.set("projects", projects);
                     }
+                    JsonNode projectNode = projects.get(cwd);
+                    ObjectNode project = projectNode instanceof ObjectNode projectObject
+                            ? projectObject : TRUST_JSON.createObjectNode();
+                    if (!(projectNode instanceof ObjectNode)) {
+                        projects.set(cwd, project);
+                    }
+                    // fleetd #247: ONLY hasTrustDialogAccepted. We used to write
+                    // hasCompletedProjectOnboarding beside it; do not put it back. Measured on
+                    // 2026-09-03, minutes after a live spawn seeded this file: 28 of 28 project
+                    // entries carried hasTrustDialogAccepted and 0 of 28 carried the onboarding key
+                    // — including the 27 entries Claude Code wrote for itself. Claude Code
+                    // normalises the whole file when it saves and drops that key every time, so
+                    // writing it achieved nothing except making the next reader think it mattered.
+                    // The member reached idle with the trust flag alone, which is the only outcome
+                    // this seed exists for. If a future Claude Code needs the second flag the
+                    // symptom returns as the trust dialog fleetd #149 describes — re-measure then,
+                    // do not restore it on a guess.
+                    project.put("hasTrustDialogAccepted", true);
+                    String newContent = TRUST_JSON.writerWithDefaultPrettyPrinter().writeValueAsString(root);
+
+                    trustJsonCasTestHook.run();
+
+                    // fleetd #247 CAS: re-read immediately before the move and compare with what
+                    // this attempt built its update from. A mismatch means another writer (most
+                    // plausibly the operator's own live Claude Code — see this method's javadoc)
+                    // landed a change in between; discard this attempt's work and rebuild from the
+                    // fresh bytes rather than blindly overwriting it.
+                    byte[] atMove = Files.isRegularFile(target) ? Files.readAllBytes(target) : null;
+                    if (!Arrays.equals(before, atMove)) {
+                        continue;
+                    }
+                    writeAtomically(target, newContent);
+                    return;
                 }
-                if (root == null) {
-                    root = TRUST_JSON.createObjectNode();
-                }
-                JsonNode projectsNode = root.get("projects");
-                ObjectNode projects = projectsNode instanceof ObjectNode projectsObject
-                        ? projectsObject : TRUST_JSON.createObjectNode();
-                if (!(projectsNode instanceof ObjectNode)) {
-                    root.set("projects", projects);
-                }
-                JsonNode projectNode = projects.get(cwd);
-                ObjectNode project = projectNode instanceof ObjectNode projectObject
-                        ? projectObject : TRUST_JSON.createObjectNode();
-                if (!(projectNode instanceof ObjectNode)) {
-                    projects.set(cwd, project);
-                }
-                // fleetd #247: ONLY hasTrustDialogAccepted. We used to write
-                // hasCompletedProjectOnboarding beside it; do not put it back. Measured on
-                // 2026-09-03, minutes after a live spawn seeded this file: 28 of 28 project
-                // entries carried hasTrustDialogAccepted and 0 of 28 carried the onboarding key
-                // — including the 27 entries Claude Code wrote for itself. Claude Code
-                // normalises the whole file when it saves and drops that key every time, so
-                // writing it achieved nothing except making the next reader think it mattered.
-                // The member reached idle with the trust flag alone, which is the only outcome
-                // this seed exists for. If a future Claude Code needs the second flag the
-                // symptom returns as the trust dialog fleetd #149 describes — re-measure then,
-                // do not restore it on a guess.
-                project.put("hasTrustDialogAccepted", true);
-                writeAtomically(target, TRUST_JSON.writerWithDefaultPrettyPrinter().writeValueAsString(root));
+                // fleetd #247: deliberately do NOT write here. A member that starts without the
+                // seed still starts — it may hit the trust dialog fleetd #149 describes and fail to
+                // reach an injectable state, but that failure is visible (herdr reports it, the
+                // spawn-readiness gate times out) and recoverable (retry the spawn). Writing our
+                // stale copy over whatever the other writer left would be silent and, if that other
+                // writer is the operator's own live session, could destroy real configuration —
+                // fail toward the recoverable outcome, not the silent one.
+                log.warn("seedTrustDialog: gave up seeding workspace-trust for cwd '{}' into '{}' "
+                        + "after {} attempts — another writer (most plausibly the operator's own "
+                        + "live Claude Code sharing this file) kept changing it faster than we could "
+                        + "re-read it, so nothing was written; the member may show the trust dialog "
+                        + "instead", cwd, target, MAX_TRUST_JSON_CAS_ATTEMPTS);
             } catch (Exception e) {
                 log.debug("cannot seed workspace-trust entry for cwd '{}' into '{}'", cwd, target, e);
             }
         }
     }
 
+    /** Bound on {@link #seedTrustDialog}'s fleetd #247 compare-and-swap retry loop. */
+    private static final int MAX_TRUST_JSON_CAS_ATTEMPTS = 5;
+
+    /**
+     * Test-only seam for fleetd #247: invoked once per CAS attempt inside {@link #seedTrustDialog}'s
+     * retry loop, after that attempt has read the target's bytes and built its replacement content,
+     * but immediately before the final re-read/compare that decides whether to write. A no-op in
+     * production. Package-visible (not {@code private}) so {@code ClaudeCodeLauncherTest} can install
+     * a hook here that writes to the target file, deterministically simulating a writer racing
+     * fleetd's own read-modify-write at the exact instant the CAS is meant to catch — the same race
+     * an external process (most plausibly the operator's own live Claude Code) creates, without
+     * depending on real thread scheduling to land the interleaving. A test that sets this MUST
+     * restore it to the no-op default in a {@code finally} block — it is shared, static state.
+     */
+    static Runnable trustJsonCasTestHook = () -> {};
+
+    /**
+     * Parse {@code bytes} as a {@code .claude.json} tree, or hand back a fresh empty object when
+     * {@code bytes} is {@code null} (no file yet) or does not parse to a JSON object — the same
+     * missing-or-unreadable-is-empty fallback {@link #seedTrustDialog} always used, factored out so
+     * the fleetd #247 CAS loop can call it once per attempt.
+     */
+    private static ObjectNode parseTrustJsonOrEmpty(byte[] bytes) throws IOException {
+        if (bytes == null) {
+            return TRUST_JSON.createObjectNode();
+        }
+        JsonNode existing = TRUST_JSON.readTree(bytes);
+        return existing instanceof ObjectNode existingObject ? existingObject : TRUST_JSON.createObjectNode();
+    }
+
     /**
      * Write {@code content} to {@code target} atomically: serialise to a sibling temp file in the
      * <strong>same directory</strong> as {@code target} (an atomic move is only guaranteed within

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

@@ -117,9 +117,11 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      *
      * <p>fleetd #185 stage 2: that mirroring assumption holds only while the member pane runs under
      * the SAME OS user as the daemon. When {@code memberHerdrSocket:} is configured, member panes
-     * run on a second herdr owned by a different user — different {@code $HOME}, different {@code
-     * secrets.sh}, different environment entirely — so this field's data no longer describes what a
-     * member pane inherits. See {@link #logCredentialGap} for how that mode is handled.
+     * are routed to a second herdr, and fleetd has no channel to confirm what OS user that herdr
+     * runs as — it may be a different user with a different {@code $HOME} and {@code secrets.sh},
+     * or the same one the daemon runs as. Either way this field's data can no longer be trusted to
+     * describe what a member pane inherits. See {@link #logCredentialGap} for how that mode is
+     * handled.
      */
     private final Supplier<Set<String>> hostEnvNames;
 
@@ -1471,9 +1473,9 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         Set<String> brokerUriEnvNames = brokerUriEnvNames();
         Set<String> allowed = new java.util.TreeSet<>(
                 MemberEnvAllowList.derive(profiles.values(), creds.allowSet(), brokerUriEnvNames));
-        if (creds.sshAuthSockAllowed()) {
+        if (creds.sshAgentEnvInherited()) {
             allowed.add(SSH_AUTH_SOCK);
-        } // blocked by default: absent from the set ⇒ blanked by the scrub like any other name
+        } // omitted by default: absent from the set ⇒ blanked by the scrub like any other name
         allowed.addAll(launch.env().keySet());
         allowed.removeAll(brokerUriEnvNames);
         return allowed;
@@ -1533,21 +1535,24 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
     /**
      * fleetd #213: {@code memberHerdrSocket} is configured and the member login shell IS zsh, but
      * {@code worktreeRoot} and/or {@code worktreeGroup} is missing, so the generated ZDOTDIR cannot
-     * be placed anywhere the member's OS user can reach — {@code java.io.tmpdir} is fleetd's own
-     * 0700 temp dir, unreadable by another uid, which is the exact gap this ticket exists to close.
-     * Say so once per launcher instance, instead of either generating a directory nothing can read
-     * (protection theatre) or refusing to spawn (turning a degraded credential control into an
-     * outage for an opt-in feature).
+     * be placed anywhere fleetd can be sure the member's OS user can reach — {@code java.io.tmpdir}
+     * is fleetd's own 0700 temp dir, which is unreadable if the member pane runs as a different OS
+     * user, and fleetd has no channel to confirm whether it does or not. Rather than gamble on that,
+     * this treats memberHerdrSocket as reason enough to require an explicitly shared location, which
+     * is the exact gap this ticket exists to close. Say so once per launcher instance, instead of
+     * either generating a directory that might not be readable (protection theatre) or refusing to
+     * spawn (turning a degraded credential control into an outage for an opt-in feature).
      */
     private void warnCannotShareScrubDirectory() {
         if (cannotShareScrubDirWarned.compareAndSet(false, true)) {
             log.warn("memberCredentials policy=allow-list: memberHerdrSocket is configured and the "
                     + "member login shell is zsh, but worktreeRoot and/or worktreeGroup is not "
-                    + "configured — the generated ZDOTDIR cannot be placed where the member's OS "
-                    + "user can read it (java.io.tmpdir is fleetd's own, unreadable by another uid), "
-                    + "so the scrub cannot be guaranteed to run. Falling back to the CB-596 sentinel "
-                    + "overlay. Configure both worktreeRoot and worktreeGroup to enable the "
-                    + "allow-list scrub under memberHerdrSocket.");
+                    + "configured — the generated ZDOTDIR cannot be placed where fleetd can be sure "
+                    + "the member's OS user can read it (java.io.tmpdir is fleetd's own 0700 dir, "
+                    + "unreadable if the member runs as a different OS user — fleetd has no channel "
+                    + "to confirm whether it does), so the scrub cannot be guaranteed to run. Falling "
+                    + "back to the CB-596 sentinel overlay. Configure both worktreeRoot and "
+                    + "worktreeGroup to enable the allow-list scrub under memberHerdrSocket.");
         }
     }
 
@@ -1626,10 +1631,12 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
     private final AtomicBoolean unknownMemberEnvironmentWarned = new AtomicBoolean();
 
     /**
-     * fleetd #185 stage 2: whether {@code memberHerdrSocket:} is configured, i.e. member panes run
-     * on a second herdr owned by a different OS user than the daemon's own process. Re-read from the
-     * live config on every call (same hot-reload shape as {@link #memberCredentials}), never cached,
-     * so a config reload takes effect on the next spawn without a restart.
+     * fleetd #185 stage 2: whether {@code memberHerdrSocket:} is configured, i.e. member panes are
+     * routed to a second herdr. This tests only that the config key is set — fleetd has no channel
+     * to confirm what OS user that second herdr runs as, so a {@code true} result means "member
+     * panes may run under a different OS user," not that they do. Re-read from the live config on
+     * every call (same hot-reload shape as {@link #memberCredentials}), never cached, so a config
+     * reload takes effect on the next spawn without a restart.
      *
      * <p>{@link #config} is {@code null} on any call site that never threaded the full config
      * through (every production {@code HerdrPeerLauncher} does; a handful of older tests do not) —
@@ -1652,14 +1659,15 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
      * fleetd #185 stage 2: the single replacement WARN for {@link #logCredentialGap}'s usual
      * conclusions when {@code memberHerdrSocket:} is configured. {@link #hostEnvNames} (and
      * everything derived from it — {@code known}/{@code allow} coverage, the allow-list scrub's
-     * derived set) describes the DAEMON's own environment; under this config key member panes run as
-     * a different OS user with a different environment entirely, so neither "every member pane
-     * inherits them UNBLOCKED" nor "the scrub blanks them" is evidence-backed here — both would be
-     * reporting on the wrong process. Logged once, names the config key, and states the honest
-     * conclusion: the gap for member panes is UNKNOWN, not clean, so {@code memberCredentials} cannot
-     * be verified from this daemon. The one count it does report is scoped explicitly to fleetd's own
-     * environment, never presented as if it said anything about the member's — see {@link
-     * #logCredentialGap}'s javadoc for why this branch exists.
+     * derived set) describes the DAEMON's own environment; under this config key member panes are
+     * routed to a second herdr, and fleetd has no channel to confirm what OS user that herdr runs
+     * as or to read its environment, so neither "every member pane inherits them UNBLOCKED" nor
+     * "the scrub blanks them" is evidence-backed here — both would be reporting on the wrong
+     * process. Logged once, names the config key, and states the honest conclusion: the gap for
+     * member panes is UNKNOWN, not clean, so {@code memberCredentials} cannot be verified from this
+     * daemon. The one count it does report is scoped explicitly to fleetd's own environment, never
+     * presented as if it said anything about the member's — see {@link #logCredentialGap}'s javadoc
+     * for why this branch exists.
      */
     private void warnUnknownMemberEnvironment(FleetConfig.MemberCredentials creds) {
         if (!unknownMemberEnvironmentWarned.compareAndSet(false, true)) {
@@ -1672,13 +1680,13 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
                 .filter(name -> CREDENTIAL_SHAPED_NAME.matcher(name).matches())
                 .filter(name -> !covered.contains(name))
                 .count();
-        log.warn("memberCredentials gap: memberHerdrSocket is configured, so member panes run under "
-                        + "a different OS user than fleetd's own process, with a different environment "
-                        + "entirely — fleetd has no channel to read that user's environment. {} of the "
-                        + "{} names in fleetd's OWN environment are credential-shaped and not on "
-                        + "known:/allow:, but that count describes fleetd's process, not the member "
-                        + "herdr's. The credential gap for member panes is UNKNOWN, not clean, and "
-                        + "memberCredentials cannot be verified from here.",
+        log.warn("memberCredentials gap: memberHerdrSocket is configured, so member panes are routed "
+                        + "to a second herdr — fleetd has no channel to confirm what OS user that herdr "
+                        + "runs as, so it cannot tell whether those panes inherit its own environment or "
+                        + "a different one entirely. {} of the {} names in fleetd's OWN environment are "
+                        + "credential-shaped and not on known:/allow:, but that count describes fleetd's "
+                        + "process, not the member herdr's. The credential gap for member panes is "
+                        + "UNKNOWN, not clean, and memberCredentials cannot be verified from here.",
                 gapInFleetdsOwnEnv, hostNames.size());
     }
 

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

@@ -34,7 +34,7 @@ import java.util.TreeSet;
  *
  * <p>{@code SSH_AUTH_SOCK} is deliberately NOT here. It is a handle to the operator's ssh-agent — a
  * member holding it can sign with the operator's keys — so keeping it is a config decision
- * ({@code memberCredentials.sshAuthSock: allow}), not a derivation default.
+ * ({@code memberCredentials.sshAgentEnv: inherit}), not a derivation default.
  *
  * <p><b>CB-633 follow-up:</b> the union also includes {@code memberCredentials.allow:} — the
  * operator's own explicit list. Before this, {@code policy: allow-list} silently ignored every name
@@ -42,7 +42,7 @@ import java.util.TreeSet;
  * turning the policy on could blank credentials working members already depended on. {@code
  * SSH_AUTH_SOCK} and configured broker URI environment names are exceptions: even when the operator
  * lists them under {@code allow:}, they are excluded here. {@code SSH_AUTH_SOCK} is added back ONLY
- * by the caller when {@code sshAuthSock: allow} is explicitly set
+ * by the caller when {@code sshAgentEnv: inherit} is explicitly set
  * (see {@link #SSH_AUTH_SOCK}'s javadoc) — it is a live handle to the operator's own ssh-agent, not
  * a value, so treating it like any other allow-listed name would hand a member every key the
  * operator's agent holds the moment they typed the name under {@code allow:} for an unrelated
@@ -53,7 +53,7 @@ public final class MemberEnvAllowList {
     /**
      * The operator's ssh-agent socket path. Deliberately excluded from {@link #derive}'s union of
      * {@code memberCredentials.allow:} — see the class javadoc's CB-633 follow-up note. Governed
-     * ONLY by {@code memberCredentials.sshAuthSock}, never by appearing in {@code allow:}.
+     * ONLY by {@code memberCredentials.sshAgentEnv}, never by appearing in {@code allow:}.
      */
     public static final String SSH_AUTH_SOCK = "SSH_AUTH_SOCK";
 

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

@@ -21,6 +21,7 @@ import java.util.EnumSet;
 import java.util.List;
 import java.util.Map;
 import java.util.Set;
+import java.util.concurrent.ConcurrentHashMap;
 import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.BooleanSupplier;
@@ -669,6 +670,16 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
     private final AtomicBoolean discoveryUnavailableWarned =
             new AtomicBoolean();
 
+    /**
+     * fleetd #267: one WARN per PROFILE (not per launcher instance — several profiles can each hit
+     * this gap independently) for the model-mismatch check (fleetd #175) never getting to run
+     * because the spawn was not given a fleetd-provisioned worktree (fleetd #249). Profile names
+     * accumulate here for the life of this launcher instance and are never removed — the same
+     * one-shot treatment {@link #discoveryUnavailableWarned} already gets, just keyed per profile
+     * instead of globally.
+     */
+    private final Set<String> modelCheckSkippedWarned = ConcurrentHashMap.newKeySet();
+
     /** Add lazy on-disk session discovery to the base handle. */
     @Override
     public PeerHandle spawn(SpawnRequest req) {
@@ -695,7 +706,8 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
         // be running.
         FleetConfig.Profile cfg = requireProfile(req.profileName());
         return new SessionAwareHandle(inner, discovery, cwd, cfg,
-                this::memberHerdrSocketConfigured, discoveryUnavailableWarned, exhaustionSink);
+                this::memberHerdrSocketConfigured, discoveryUnavailableWarned,
+                modelCheckSkippedWarned, exhaustionSink);
     }
 
     /**
@@ -720,6 +732,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
         private final FleetConfig.Profile cfg;
         private final BooleanSupplier discoveryUnavailable;
         private final AtomicBoolean discoveryUnavailableWarned;
+        private final Set<String> modelCheckSkippedWarned;
         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();
@@ -750,6 +763,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                             FleetConfig.Profile cfg,
                             BooleanSupplier discoveryUnavailable,
                             AtomicBoolean discoveryUnavailableWarned,
+                            Set<String> modelCheckSkippedWarned,
                             ExhaustionSink exhaustionSink) {
             this.delegate = delegate;
             this.discovery = discovery;
@@ -757,6 +771,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
             this.cfg = cfg;
             this.discoveryUnavailable = discoveryUnavailable;
             this.discoveryUnavailableWarned = discoveryUnavailableWarned;
+            this.modelCheckSkippedWarned = modelCheckSkippedWarned;
             this.exhaustionSink = exhaustionSink;
             this.worktreeProvisioned = isProvisionedWorktree(cwd);
         }
@@ -807,9 +822,29 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
             // member's row apart from a sibling's in that case (measured: a three-day-old row from
             // a different profile). Refuse to guess — absent is the honest answer, and it is what
             // this codebase already returns elsewhere for absent evidence (fleetd #175's UNKNOWN).
-            // No WARN here: unlike discoveryUnavailable above, this is the ordinary, expected shape
-            // of the large majority of spawns (no worktree requested), not a configuration gap.
+            // This IS the ordinary, expected shape of the large majority of spawns (no worktree
+            // requested), not a configuration gap — but fleetd #267 found that same shape silently
+            // switches off the fleetd #175 model-mismatch check for those spawns too, since
+            // checkModelMatch's only call site is right below this gate. The check cannot be moved
+            // off agentSessionId()'s resolved id: the id is the only safe way to key
+            // actualModelForSessionId to THIS session's own row rather than "whatever is newest in
+            // the shared directory" (fleetd #234) — re-deriving a second, independent answer via
+            // `directory` here would reintroduce exactly the false-positive risk #234 fixed (a
+            // sibling's differently-configured model looking like THIS profile's mismatch). So the
+            // model genuinely is unknowable without a provisioned worktree, and unlike the silence
+            // this branch used to keep, that gap now gets the same one-time, per-profile WARN
+            // treatment discoveryUnavailable already gets above — but keyed by profile, since
+            // several profiles can each hit this independently.
             if (!worktreeProvisioned) {
+                if (cfg.model() != null && !cfg.model().isBlank()
+                        && modelCheckSkippedWarned.add(cfg.profile())) {
+                    log.warn("opencode model-mismatch check (fleetd #175) cannot run for profile "
+                            + "'{}': it was spawned without a fleetd-provisioned worktree (fleetd "
+                            + "#249), so its cwd may be shared with other sessions and the actual "
+                            + "model it is running cannot be safely told apart from a sibling's — "
+                            + "spawn with worktree:true to enable the check for this profile.",
+                            cfg.profile());
+                }
                 return null;
             }
             // fleetd #234: once resolved, stay resolved. Re-deriving from `directory` on every call

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

@@ -0,0 +1,73 @@
+package dev.ltms.fleet;
+
+import ch.qos.logback.classic.Logger;
+import ch.qos.logback.classic.spi.ILoggingEvent;
+import ch.qos.logback.core.read.ListAppender;
+import dev.ltms.fleet.config.FleetConfig;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.io.TempDir;
+import org.slf4j.LoggerFactory;
+
+import java.nio.file.Files;
+import java.nio.file.Path;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+
+/**
+ * fleetd #184: startup must state whether members share fleetd's OS user or use a separate herdr.
+ */
+class MemberTrustModelReportTest {
+
+    private static FleetConfig load(Path dir, String yaml) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml);
+        return FleetConfig.load(f);
+    }
+
+    private static ListAppender<ILoggingEvent> attach() {
+        Logger logger = (Logger) LoggerFactory.getLogger(Fleetd.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        return appender;
+    }
+
+    private static void detach(ListAppender<ILoggingEvent> appender) {
+        ((Logger) LoggerFactory.getLogger(Fleetd.class)).detachAppender(appender);
+    }
+
+    private static String report(FleetConfig cfg) {
+        Logger logger = (Logger) LoggerFactory.getLogger(Fleetd.class);
+        ch.qos.logback.classic.Level original = logger.getLevel();
+        logger.setLevel(ch.qos.logback.classic.Level.INFO);
+        ListAppender<ILoggingEvent> appender = attach();
+        try {
+            Fleetd.reportMemberTrustModel(cfg);
+        } finally {
+            detach(appender);
+            logger.setLevel(original);
+        }
+        return appender.list.getFirst().getFormattedMessage();
+    }
+
+    @Test
+    void unsetMemberHerdrSocketStatesThatMembersAreNotSandboxed(@TempDir Path dir) throws Exception {
+        FleetConfig cfg = load(dir, "bind:\n  host: 127.0.0.1\n  port: 8765\n");
+
+        assertEquals("member trust model: members run as the same OS user as fleetd, not in a sandbox. "
+                        + "A member can read any file this user can read, including SSH keys and credential "
+                        + "stores, whatever memberCredentials says. To add a real boundary, route members to "
+                        + "a second herdr under a different OS user with memberHerdrSocket.",
+                report(cfg));
+    }
+
+    @Test
+    void configuredMemberHerdrSocketStatesThatFleetdCannotConfirmTheBoundary(@TempDir Path dir) throws Exception {
+        FleetConfig cfg = load(dir, "memberHerdrSocket: /tmp/member-herdr.sock\n");
+
+        assertEquals("member trust model: members are routed to a separate herdr through "
+                        + "memberHerdrSocket. fleetd cannot see that herdr's uid, so confirm it runs "
+                        + "as a different OS user before treating it as a boundary.",
+                report(cfg));
+    }
+}

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

@@ -0,0 +1,96 @@
+package dev.ltms.fleet;
+
+import com.tngtech.archunit.base.DescribedPredicate;
+import com.tngtech.archunit.core.domain.JavaClass;
+import com.tngtech.archunit.core.domain.JavaClass.Predicates;
+import com.tngtech.archunit.core.importer.ClassFileImporter;
+import com.tngtech.archunit.core.importer.ImportOption;
+import com.tngtech.archunit.library.dependencies.SliceRule;
+import com.tngtech.archunit.library.dependencies.SlicesRuleDefinition;
+import org.junit.jupiter.api.Test;
+
+/**
+ * fleetd #131 (CB-627): enforce package boundaries with an ArchUnit test instead of a
+ * Maven module split.
+ *
+ * <p>This test fails the build the moment a NEW cycle appears between the top-level
+ * {@code dev.ltms.fleet.*} packages. Today's cycles are recorded below as explicit,
+ * narrow exceptions: each one ignores dependencies between exactly the two named
+ * packages, in both directions, and nothing else. A cycle through any other pair of
+ * packages -- or a brand new pair -- still fails this test.
+ *
+ * <p><b>Main code only.</b> The import excludes test classes
+ * ({@link ImportOption.Predefined#DO_NOT_INCLUDE_TESTS}). Test code legitimately wires
+ * across many packages for setup and mocking; that is not part of the shipped
+ * architecture this rule protects. Verified: importing test classes too pulls in a much
+ * larger, noisier cycle set -- {@code herdr}, {@code member}, {@code peer}, {@code
+ * config}, {@code guard} and {@code placement} all show up in cycles that disappear the
+ * moment test classes are excluded. Scanning off the classpath via {@code
+ * importPackages(...)} (not a hardcoded {@code target/classes} path) also keeps this
+ * test correct regardless of the working directory the build is invoked from.
+ *
+ * <p><b>No package moves here</b> -- ticket #131 is explicit that removing a cycle is
+ * its own, later PR. See the comment on each exception below for which ticket step
+ * removes it.
+ */
+class PackageCyclesTest {
+
+    @Test
+    void packagesAreFreeOfCycles() {
+        var classes = new ClassFileImporter()
+                .withImportOption(ImportOption.Predefined.DO_NOT_INCLUDE_TESTS)
+                .importPackages("dev.ltms.fleet");
+
+        SliceRule rule = SlicesRuleDefinition.slices()
+                .matching("dev.ltms.fleet.(*)..")
+                .should().beFreeOfCycles();
+
+        // fleetd #131 step 1: move ConnectionIdentity so authz stops depending on the
+        // MCP layer. Evidence: auth/CallerResolver.java:3 imports mcp.ConnectionIdentity;
+        // mcp/FleetMcp.java:3-7 imports auth.AuditLog, Authz, CallerResolver, Principal,
+        // Role.
+        rule = ignoreCycle(rule, "auth", "mcp");
+
+        // fleetd #131 step 2: PrimaryRegistry is used by loops in msg; move it, or put
+        // an interface between msg and mcp. Evidence: msg/ReplyPushLoop.java:5 and
+        // msg/LeadHeartbeatLoop.java:5 import mcp.PrimaryRegistry; mcp/FleetMcp.java:15-18
+        // imports msg.LeadChannel, LeadMessage, MessageService, Rendezvous.
+        rule = ignoreCycle(rule, "mcp", "msg");
+
+        // fleetd #131 -- found while implementing this test, NOT one of the ticket's
+        // original three; it names its own follow-up step before removal. Evidence:
+        // inject/CompletionResolver.java:4-5, inject/Injector.java:6 and
+        // inject/TurnListener.java:3 import msg.Rendezvous / msg.TurnToken;
+        // msg/MessageService.java:6 imports inject.Injector.
+        rule = ignoreCycle(rule, "inject", "msg");
+
+        // fleetd #131 -- same as above, its own follow-up. Evidence:
+        // metrics/FleetMetrics.java:3 imports msg.ReplyInbox; msg/MessageService.java:7-8,
+        // msg/LeadHeartbeatLoop.java:6-7 and msg/ReplyPushLoop.java:6-7 import
+        // metrics.FleetMetrics / metrics.Metrics.
+        rule = ignoreCycle(rule, "metrics", "msg");
+
+        // fleetd #131 -- same as above, its own follow-up. Evidence:
+        // session/SessionManager.java:7 imports msg.TurnToken;
+        // msg/LeadHeartbeatLoop.java:8 imports session.MemberSession.
+        rule = ignoreCycle(rule, "msg", "session");
+
+        rule.check(classes);
+    }
+
+    /**
+     * Accepts today's known cycle between two top-level packages, and nothing else.
+     * Ignoring both directions removes exactly this pair from cycle detection; every
+     * other dependency -- including any new one added later, between these same two
+     * packages or any other pair -- is still checked.
+     */
+    private static SliceRule ignoreCycle(SliceRule rule, String packageA, String packageB) {
+        return rule
+                .ignoreDependency(residesIn(packageA), residesIn(packageB))
+                .ignoreDependency(residesIn(packageB), residesIn(packageA));
+    }
+
+    private static DescribedPredicate<JavaClass> residesIn(String topLevelPackage) {
+        return Predicates.resideInAPackage("dev.ltms.fleet." + topLevelPackage + "..");
+    }
+}

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

@@ -2010,22 +2010,93 @@ class FleetConfigTest {
                 "deny-list normalizes onto the canonical deny-by-default value");
     }
 
-    /**
-     * CB-633: {@code SSH_AUTH_SOCK} is a decision, never a default — absent, blank, or misspelled,
-     * it stays BLOCKED; only the literal "allow" (any case) passes it through. A typo like "alow"
-     * failing safe here is the whole point of making it a knob.
-     */
     @Test
-    void sshAuthSockDefaultsToBlockedAndOnlyExplicitAllowUnblocksIt() {
-        assertTrue(new FleetConfig.MemberCredentials("allow-list", List.of(), List.of()).sshAuthSock().equals("block"),
-                "absent knob blocks SSH_AUTH_SOCK");
-        assertFalse(new FleetConfig.MemberCredentials("allow-list", List.of(), List.of()).sshAuthSockAllowed());
-        assertFalse(new FleetConfig.MemberCredentials(null, null, null, "").sshAuthSockAllowed(),
-                "blank knob blocks SSH_AUTH_SOCK");
-        assertFalse(new FleetConfig.MemberCredentials(null, null, null, "alow").sshAuthSockAllowed(),
-                "a misspelled value fails SAFE, not open");
-        assertTrue(new FleetConfig.MemberCredentials(null, null, null, "ALLOW").sshAuthSockAllowed(),
-                "the literal allow (case-insensitive) unblocks SSH_AUTH_SOCK");
+    void sshAgentEnvAcceptsEveryCompatibleKeyAndValuePair(@TempDir Path dir) throws Exception {
+        String[][] spellings = {
+                {"sshAuthSock", "block", "omit"},
+                {"sshAuthSock", "allow", "inherit"},
+                {"sshAuthSock", "omit", "omit"},
+                {"sshAuthSock", "inherit", "inherit"},
+                {"sshAgentEnv", "block", "omit"},
+                {"sshAgentEnv", "allow", "inherit"},
+                {"sshAgentEnv", "omit", "omit"},
+                {"sshAgentEnv", "inherit", "inherit"}
+        };
+
+        for (int i = 0; i < spellings.length; i++) {
+            Path file = dir.resolve("member-credentials-ssh-agent-" + i + ".yaml");
+            Files.writeString(file, """
+                    bind:
+                      port: 8080
+                    memberCredentials:
+                      policy: allow-list
+                    """ + "  " + spellings[i][0] + ": " + spellings[i][1] + "\n");
+
+            FleetConfig.MemberCredentials credentials = FleetConfig.load(file).memberCredentials();
+            assertEquals(spellings[i][2], credentials.sshAgentEnv(),
+                    spellings[i][0] + ": " + spellings[i][1] + " must normalize correctly");
+            assertEquals("inherit".equals(spellings[i][2]), credentials.sshAgentEnvInherited());
+        }
+    }
+
+    /** Protects the live {@code sshAuthSock: block} allow-list configuration during the rename. */
+    @Test
+    void legacySshAuthSockBlockKeepsLiveAllowListConfigOmitted(@TempDir Path dir) throws Exception {
+        Path file = dir.resolve("live-member-credentials.yaml");
+        Files.writeString(file, """
+                bind:
+                  port: 8080
+                memberCredentials:
+                  policy: allow-list
+                  sshAuthSock: block
+                """);
+
+        FleetConfig.MemberCredentials credentials = FleetConfig.load(file).memberCredentials();
+
+        assertEquals("omit", credentials.sshAgentEnv());
+        assertFalse(credentials.sshAgentEnvInherited(), "the live config must omit SSH_AUTH_SOCK");
+    }
+
+    @Test
+    void sshAgentEnvWinsWhenBothCompatibleKeysArePresent(@TempDir Path dir) throws Exception {
+        Path file = dir.resolve("both-ssh-agent-keys.yaml");
+        Files.writeString(file, """
+                bind:
+                  port: 8080
+                memberCredentials:
+                  policy: allow-list
+                  sshAuthSock: allow
+                  sshAgentEnv: omit
+                """);
+
+        FleetConfig.MemberCredentials credentials = FleetConfig.load(file).memberCredentials();
+
+        assertEquals("omit", credentials.sshAgentEnv());
+        assertFalse(credentials.sshAgentEnvInherited());
+    }
+
+    @Test
+    void sshAgentEnvDefaultsToOmitAndUnknownValuesFailClosed(@TempDir Path dir) throws Exception {
+        Path absent = dir.resolve("member-credentials-ssh-agent-absent.yaml");
+        Files.writeString(absent, """
+                bind:
+                  port: 8080
+                memberCredentials:
+                  policy: allow-list
+                """);
+        assertEquals("omit", FleetConfig.load(absent).memberCredentials().sshAgentEnv());
+
+        Path unknown = dir.resolve("member-credentials-ssh-agent-unknown.yaml");
+        Files.writeString(unknown, """
+                bind:
+                  port: 8080
+                memberCredentials:
+                  policy: allow-list
+                  sshAgentEnv: inhert
+                """);
+        FleetConfig.MemberCredentials credentials = FleetConfig.load(unknown).memberCredentials();
+        assertEquals("omit", credentials.sshAgentEnv());
+        assertFalse(credentials.sshAgentEnvInherited(), "an unknown value must fail closed");
     }
 
     /**

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

@@ -34,6 +34,7 @@ import java.util.Map;
 import java.util.Set;
 import java.util.concurrent.CompletableFuture;
 import java.util.concurrent.TimeUnit;
+import java.util.function.Function;
 
 import static org.junit.jupiter.api.Assertions.*;
 
@@ -810,13 +811,16 @@ class FleetMcpTest {
     }
 
     /**
-     * fleetd #176: this is the exact shape measured on the Mac fleet — {@code maxLoad:3, live:2},
-     * where one of the "free" three is really the lead's own seat on the same subscription. The old
-     * formula ({@code max(0, cap - live)}) reported {@code free:1}; the real ceiling is {@code 0}
-     * (two members plus the lead's own seat already fill all three).
+     * fleetd #257: this is the exact shape measured on the Mac fleet — {@code maxLoad:3, live:2},
+     * one lead sharing the subscription. The OLD formula ({@code max(0, cap - live - leadSeats)})
+     * reported {@code free:0} here, disagreeing with the real spawn gate (which never read
+     * {@code leadSeats} and would still grant one more spawn — see
+     * {@code freeMatchesWhatTheRealPlacementGateActuallyGrants} below for that proof against the
+     * actual gate). {@code free} must report {@code 1}: {@code leadSeats} is carried as a fact, not
+     * subtracted.
      */
     @Test
-    void leadSeatSubtractsFromFreeTheSameWayLiveDoes() {
+    void leadSeatIsReportedButNeverSubtractedFromFree() {
         FakeHerdr h = new FakeHerdr();
         SessionManager sessions = new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")));
         FleetMcp.LeadSeatSource leadSeats = new FleetMcp.LeadSeatSource(
@@ -829,17 +833,17 @@ class FleetMcpTest {
 
         assertTrue(out.contains("\"maxLoad\":3"), "maxLoad itself must be left untouched: " + out);
         assertTrue(out.contains("\"live\":2"), out);
-        assertTrue(out.contains("\"free\":0"), "2 live + 1 lead seat fills all 3: " + out);
-        assertTrue(out.contains("\"leadSeats\":1"), out);
+        assertTrue(out.contains("\"free\":1"), "free is maxLoad - live only, never minus leadSeats: " + out);
+        assertTrue(out.contains("\"leadSeats\":1"), "leadSeats is still reported, just not subtracted: " + out);
     }
 
     /**
-     * fleetd #176: the OTHER measurement in the issue — a completely idle fleet still overstates
-     * {@code free} by the lead's own seat. {@code maxLoad:3, live:0} must report {@code free:2}, the
-     * real fan-out ceiling, not {@code 3}.
+     * fleetd #257: the OTHER measurement in the issue — a completely idle fleet with a lead sharing
+     * the subscription. {@code maxLoad:3, live:0} must report {@code free:3}, matching what the
+     * spawn gate (which has no notion of a lead's seat) would actually grant.
      */
     @Test
-    void leadSeatLowersFreeOnAnOtherwiseIdleSubscriptionProfile() {
+    void leadSeatDoesNotLowerFreeOnAnOtherwiseIdleSubscriptionProfile() {
         FakeHerdr h = new FakeHerdr();
         SessionManager sessions = new SessionManager(workerService(h, "http://gx00.gw:8000", Set.of("gx00.gw")));
         FleetMcp.LeadSeatSource leadSeats = new FleetMcp.LeadSeatSource(
@@ -851,10 +855,76 @@ class FleetMcpTest {
                 FleetMcp.QuarantineSource.none(), FleetMcp.OutageSource.none(), leadSeats, Map.of(), "", null));
 
         assertTrue(out.contains("\"live\":0"), out);
-        assertTrue(out.contains("\"free\":2"), "an idle fleet's real ceiling is 3 minus the lead's own seat: " + out);
+        assertTrue(out.contains("\"free\":3"), "the real gate never subtracts the lead's seat: " + out);
         assertTrue(out.contains("\"leadSeats\":1"), out);
     }
 
+    /**
+     * fleetd #257 — the defect, driven against the REAL placement gate, not a copy of its
+     * arithmetic. {@code free} must equal exactly how many more spawns
+     * {@link CompositePeerLauncher#spawn} (routed through the same {@link SessionManager} fleet_spawn
+     * itself uses) will grant on this profile right now: this test reads whatever number
+     * {@code fleet_list}'s real {@code listFleet} call reports, then drives that many spawns through
+     * the REAL composite launcher and asserts every one succeeds, and the next one — one past what
+     * fleet_list promised — is refused. A test that instead hand-computed {@code cap - live} and
+     * compared it to {@code free} would pass even if both sides shared the same wrong formula (this
+     * repo has been bitten by exactly that before); this one only passes when fleet_list's number and
+     * the gate's real behaviour actually agree.
+     *
+     * <p>{@code liveCount} here is wired the same way {@code Fleetd.main} wires it in production: one
+     * function, read by both the {@link CompositePeerLauncher}'s {@code maxLoad} gate and
+     * {@code fleet_list}'s {@code CapacitySource}, off the SAME {@link SessionManager#roster()} — so
+     * the two paths cannot silently drift on what "live" means.
+     */
+    @Test
+    void freeMatchesWhatTheRealPlacementGateActuallyGrants() {
+        FakeHerdr h = new FakeHerdr();
+        FleetConfig.Profile wcfg = new FleetConfig.Profile(
+                "sonnet", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN", null,
+                "tab", "fleetd-workers", "worker: {profile} #{n}", null,
+                null, null, null, null, null, null, null, 3);
+        Map<String, FleetConfig.Profile> profiles = Map.of(wcfg.profile(), wcfg);
+        ClaudeCodeLauncher delegate = new ClaudeCodeLauncher(
+                new AgentControl(h), new WorkspaceControl(h), new SubscriptionGuard(Set.of("gx00.gw")),
+                profiles, wcfg.profile(), k -> "FLEETD_WORKER_TOKEN".equals(k) ? "tok" : null);
+
+        java.util.concurrent.atomic.AtomicReference<SessionManager> smRef =
+                new java.util.concurrent.atomic.AtomicReference<>();
+        Function<String, Integer> liveCount = profile -> (int) smRef.get().roster().stream()
+                .filter(s -> profile.equals(s.profile())).count();
+        CompositePeerLauncher composite = new CompositePeerLauncher(
+                List.of(delegate), wcfg.profile(), profiles, PlacementPolicies.fixed(), liveCount);
+        SessionManager sm = new SessionManager(composite);
+        smRef.set(sm);
+
+        // Two members already live — the exact shape measured in fleetd #257 (maxLoad:3, live:2).
+        assertFalse(FleetMcp.spawn(sm, "sonnet").isError(), "setup: first live member must spawn cleanly");
+        assertFalse(FleetMcp.spawn(sm, "sonnet").isError(), "setup: second live member must spawn cleanly");
+
+        // A lead session shares this profile's subscription — leadSeats:1, same as the ticket.
+        FleetMcp.LeadSeatSource leadSeats = new FleetMcp.LeadSeatSource(p -> "sonnet".equals(p) ? 1 : 0);
+        String out = textOf(FleetMcp.listFleet(composite, sm, null,
+                new FleetMcp.CapacitySource(liveCount, p -> profiles.get(p).maxLoad(), profiles::keySet, () -> 0),
+                new FleetMcp.HealthCoverageSource(() -> "off"), FleetMcp.QuarantineSource.none(),
+                FleetMcp.OutageSource.none(), leadSeats, Map.of(), "", null));
+        int reportedFree = extractInt(out, "free");
+
+        for (int i = 0; i < reportedFree; i++) {
+            McpSchema.CallToolResult res = FleetMcp.spawn(sm, "sonnet");
+            assertFalse(res.isError(), "fleet_list promised free:" + reportedFree + "; spawn #" + (i + 1)
+                    + " of that many was refused by the real gate: " + textOf(res));
+        }
+        McpSchema.CallToolResult overflow = FleetMcp.spawn(sm, "sonnet");
+        assertTrue(overflow.isError(), "fleet_list reported free:" + reportedFree
+                + " but the real placement gate granted at least one more spawn than that: " + textOf(overflow));
+    }
+
+    private static int extractInt(String json, String key) {
+        java.util.regex.Matcher m = java.util.regex.Pattern.compile("\"" + key + "\":(-?\\d+)").matcher(json);
+        assertTrue(m.find(), "no \"" + key + "\" field in: " + json);
+        return Integer.parseInt(m.group(1));
+    }
+
     /** A profile with no lead seats reported must be byte-identical to before this ticket. */
     @Test
     void zeroLeadSeatsOmitsTheKeyAndLeavesFreeUnchanged() {

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

@@ -26,6 +26,7 @@ import org.junit.jupiter.api.io.TempDir;
 import org.slf4j.LoggerFactory;
 
 import java.io.IOException;
+import java.io.UncheckedIOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.nio.file.attribute.PosixFileAttributeView;
@@ -39,6 +40,7 @@ import java.util.UUID;
 import java.util.concurrent.CountDownLatch;
 import java.util.concurrent.TimeUnit;
 import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicInteger;
 import java.util.concurrent.atomic.AtomicReference;
 import java.util.function.Function;
 import java.util.function.Supplier;
@@ -2594,4 +2596,150 @@ class ClaudeCodeLauncherTest {
                 + tornRead.get());
         assertEquals(newContent, Files.readString(target), "the final content must be the new content");
     }
+
+    // --- fleetd #247: CAS against a writer TRUST_JSON_LOCK cannot reach --------------------------
+    //
+    // fleetd #149's lock only serialises seedTrustDialog calls THIS launcher makes inside this one
+    // JVM. It does nothing about the one writer that actually shares this file on a real host: the
+    // operator's own live Claude Code, whose CLAUDE_CONFIG_DIR is routinely the very configDir this
+    // profile is given. A plain read-modify-write there is a routine lost update — fleetd reads v1,
+    // the operator's session writes v2, fleetd's ATOMIC_MOVE lands v3 built from v1, and v2 is gone,
+    // atomically. The fix re-reads the file's exact bytes immediately before the move and compares
+    // them with what the update was built from, retrying from fresh bytes on a mismatch.
+    //
+    // Both tests below drive the race through the real seedTrustDialog/spawn() path (not a
+    // hand-rolled call to some extracted primitive) using trustJsonCasTestHook — a seam fired once
+    // per CAS attempt, at the exact point between the read and the final compare, so the race is
+    // deterministic instead of depending on real thread timing.
+
+    @Test
+    void seedTrustDialogRetriesAndPreservesAConcurrentExternalWritersChange(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        Path claudeJson = configDir.resolve(".claude.json");
+        Files.writeString(claudeJson, "{\"projects\":{}}");
+
+        // Fires exactly once, on the first CAS attempt — simulating the operator's own live Claude
+        // Code landing its own write to this SAME file in the gap between fleetd's read and write.
+        AtomicBoolean fired = new AtomicBoolean(false);
+        ClaudeCodeLauncher.trustJsonCasTestHook = () -> {
+            if (fired.compareAndSet(false, true)) {
+                try {
+                    Files.writeString(claudeJson,
+                            "{\"projects\":{\"/operator/own/project\":"
+                                    + "{\"hasTrustDialogAccepted\":true}}}");
+                } catch (IOException e) {
+                    throw new UncheckedIOException(e);
+                }
+            }
+        };
+        try {
+            FakeHerdr herdr = new FakeHerdr();
+            FleetConfig.Profile cfg = trustProfile(configDir.toString(), worktree.toString());
+            new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                    new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null)
+                    .spawn();
+        } finally {
+            ClaudeCodeLauncher.trustJsonCasTestHook = () -> {};
+        }
+
+        assertTrue(fired.get(), "the race hook must actually have fired during the spawn");
+        JsonNode root = new ObjectMapper().readTree(claudeJson.toFile());
+        assertTrue(root.path("projects").path("/operator/own/project")
+                        .path("hasTrustDialogAccepted").asBoolean(false),
+                "the external writer's change, landed between fleetd's read and write, must SURVIVE "
+                        + "— this is the whole point of the CAS: without it, fleetd's stale-built "
+                        + "ATOMIC_MOVE would have silently discarded it. Final file: "
+                        + Files.readString(claudeJson));
+        assertTrue(root.path("projects").path(worktree.toString())
+                        .path("hasTrustDialogAccepted").asBoolean(false),
+                "fleetd's own retry must still land its own trust entry, built from the fresh bytes");
+    }
+
+    /**
+     * Retry-exhaustion path: an external writer that changes the file on EVERY attempt (not just
+     * once) exhausts all {@code MAX_TRUST_JSON_CAS_ATTEMPTS} retries. fleetd must then write nothing
+     * at all — not even a partial/best-effort write — and log a WARN naming the file it gave up on.
+     */
+    @Test
+    void seedTrustDialogWritesNothingAndWarnsWhenCasRetriesAreExhausted(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        Path claudeJson = configDir.resolve(".claude.json");
+        Files.writeString(claudeJson, "{\"marker\":\"start\"}");
+
+        AtomicInteger hookCalls = new AtomicInteger();
+        ClaudeCodeLauncher.trustJsonCasTestHook = () -> {
+            try {
+                Files.writeString(claudeJson,
+                        "{\"marker\":\"race-" + hookCalls.incrementAndGet() + "\"}");
+            } catch (IOException e) {
+                throw new UncheckedIOException(e);
+            }
+        };
+
+        Logger logger = (Logger) LoggerFactory.getLogger(ClaudeCodeLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            FakeHerdr herdr = new FakeHerdr();
+            FleetConfig.Profile cfg = trustProfile(configDir.toString(), worktree.toString());
+            new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                    new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null)
+                    .spawn();
+        } finally {
+            logger.detachAppender(appender);
+            ClaudeCodeLauncher.trustJsonCasTestHook = () -> {};
+        }
+
+        assertEquals(5, hookCalls.get(), "the race hook must fire exactly once per CAS attempt");
+        String finalContent = Files.readString(claudeJson);
+        assertEquals("{\"marker\":\"race-" + hookCalls.get() + "\"}", finalContent,
+                "the file must be left exactly as the external writer last left it — fleetd must not "
+                        + "have written at all once retries are exhausted");
+        assertFalse(finalContent.contains("hasTrustDialogAccepted"),
+                "the trust entry must never appear — its presence would mean the CAS gave up and "
+                        + "wrote anyway instead of skipping the seed");
+
+        assertTrue(appender.list.stream().anyMatch(e -> e.getLevel() == Level.WARN
+                        && e.getFormattedMessage().contains("gave up seeding workspace-trust")
+                        && e.getFormattedMessage().contains(worktree.toString())
+                        && e.getFormattedMessage().contains(claudeJson.toString())),
+                "a WARN naming both the cwd and the file it gave up on must be logged: "
+                        + appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList());
+    }
+
+    /**
+     * The loud-default WARN (fleetd #247): a profile with no {@code configDir} targets the
+     * operator's real {@code ~/.claude.json} (redirected here to a {@code @TempDir}), and every time
+     * that happens must be logged, not just detected.
+     */
+    @Test
+    void seedTrustDialogWarnsEveryTimeItTargetsTheDefaultClaudeJson(
+            @TempDir Path fakeHome, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        String originalHome = System.getProperty("user.home");
+        System.setProperty("user.home", fakeHome.toString());
+        Logger logger = (Logger) LoggerFactory.getLogger(ClaudeCodeLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            FakeHerdr herdr = new FakeHerdr();
+            FleetConfig.Profile cfg = trustProfile(null, worktree.toString());
+            new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                    new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null)
+                    .spawn();
+        } finally {
+            logger.detachAppender(appender);
+            System.setProperty("user.home", originalHome);
+        }
+
+        assertTrue(appender.list.stream().anyMatch(e -> e.getLevel() == Level.WARN
+                        && e.getFormattedMessage().contains("no configDir set")
+                        && e.getFormattedMessage().contains(worktree.toString())),
+                "a WARN naming the cwd must fire when the profile sets no configDir: "
+                        + appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList());
+    }
 }

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

@@ -170,10 +170,10 @@ class HerdrPeerLauncherAllowListWiringTest {
     /**
      * {@code SSH_AUTH_SOCK} is a live ssh-agent handle, not a value — it must stay blocked under
      * {@code allow-list} even when the operator lists it under {@code allow:}, because {@code
-     * sshAuthSock} defaults to blocked. Governed ONLY by {@code memberCredentials.sshAuthSock}.
+     * sshAgentEnv} defaults to omit. Governed ONLY by {@code memberCredentials.sshAgentEnv}.
      */
     @Test
-    void sshAuthSockStaysBlockedEvenWhenListedInMemberCredentialsAllow() {
+    void sshAgentEnvStaysOmittedEvenWhenListedInMemberCredentialsAllow() {
         FakeHerdr herdr = new FakeHerdr();
         WiringLauncher launcher = new WiringLauncher(herdr,
                 allowListWithAllow(List.of("SSH_AUTH_SOCK")));
@@ -184,7 +184,7 @@ class HerdrPeerLauncherAllowListWiringTest {
         String scrub = readAll(dir.resolve(EnvAllowListScrub.SCRUB_FILE));
         assertFalse(scrub.contains("'SSH_AUTH_SOCK'"),
                 "SSH_AUTH_SOCK must not be on the derived allow-list just because the operator put "
-                        + "it under allow: — sshAuthSock is unset here, so it defaults to block");
+                        + "it under allow: — sshAgentEnv is unset here, so it defaults to omit");
     }
 
     @Test
@@ -387,6 +387,33 @@ class HerdrPeerLauncherAllowListWiringTest {
                 + appender.list.stream().map(ILoggingEvent::getFormattedMessage).toList());
     }
 
+    /**
+     * fleetd #184 item 5: the unknown-environment WARN must state the honest reason for the
+     * UNKNOWN conclusion — fleetd has no channel to confirm what OS user the second herdr runs
+     * as — and must NOT assert as fact that member panes run under a different OS user just
+     * because {@code memberHerdrSocket} is configured. An operator may point it at a second herdr
+     * running as the SAME user, for pane isolation alone; in that case members DO inherit fleetd's
+     * environment, and asserting otherwise would tell the operator to disregard a real, known gap.
+     */
+    @Test
+    void unknownEnvironmentWarnStatesUncertaintyNotAnAssertedDifferentUser() {
+        FakeHerdr herdr = new FakeHerdr();
+        Set<String> hostEnvNames = Set.of("FLEETD_WORKER_TOKEN", "SOME_UNKNOWN_SECRET_TOKEN");
+        WiringLauncher launcher = new WiringLauncher(herdr, allowList(), "/bin/zsh", () -> hostEnvNames,
+                () -> configWithMemberHerdrSocket("/tmp/other-user-herdr.sock", "/bin/zsh"));
+
+        List<String> messages = spawnAndCaptureLogs(launcher);
+
+        assertTrue(messages.stream().anyMatch(m -> m.contains("memberHerdrSocket")
+                        && m.contains("no channel to confirm what OS user that herdr runs as")),
+                "expected the WARN to name the actual uncertainty (no channel to confirm the "
+                        + "herdr's uid), got: " + messages);
+        assertFalse(messages.stream().anyMatch(m -> m.contains("member panes run under a different OS "
+                        + "user than fleetd's own process")),
+                "the WARN must not assert as fact that members run under a different OS user just "
+                        + "because memberHerdrSocket is configured — got: " + messages);
+    }
+
     /**
      * Hard constraint: the gap detector must never log an env var VALUE, only its NAME. {@code
      * SOME_UNKNOWN_SECRET_TOKEN} resolves to a distinctive canary value through the same {@code env}

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

@@ -109,11 +109,11 @@ class MemberEnvAllowListTest {
     /**
      * {@code SSH_AUTH_SOCK} is a live handle to the operator's ssh-agent, never a value — so it must
      * stay excluded from the derived set even when the operator lists it under {@code allow:} for an
-     * unrelated reason. It is governed ONLY by {@code memberCredentials.sshAuthSock}, applied
+     * unrelated reason. It is governed ONLY by {@code memberCredentials.sshAgentEnv}, applied
      * separately by the caller ({@code HerdrPeerLauncher}).
      */
     @Test
-    void sshAuthSockInMemberCredentialsAllowIsStillExcluded() {
+    void sshAgentEnvInMemberCredentialsAllowIsStillExcluded() {
         Set<String> derived = MemberEnvAllowList.derive(List.of(), Set.of("SSH_AUTH_SOCK", "OTHER_NAME"));
 
         assertFalse(derived.contains("SSH_AUTH_SOCK"),

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

@@ -1440,4 +1440,131 @@ class OpenCodeLauncherTest {
                 "the profile hint must survive the Fleetd-style forwarding hop and reach the real "
                         + "sink — a lambda forwarder drops it and this must go red");
     }
+
+    // --- fleetd #267: the #175 check never ran for the ordinary (no-worktree) spawn shape --------
+
+    /**
+     * fleetd #267 acceptance criterion 2, half 1 — a regression guard for the NEW code path only:
+     * a spawn WITH a fleetd-provisioned worktree must keep running the fleetd #175 model check
+     * exactly as before (already proven thoroughly above), and must now ALSO never emit the new
+     * fleetd #267 "cannot run" WARN, since the check is not skipped in this shape. Driven through
+     * the real {@code SessionManager.acquire()}/{@code get()} late-resolve path (fleetd #209),
+     * the same path the existing #175 tests already exercise.
+     */
+    @Test
+    void aProvisionedWorktreeSpawnRunsTheModelCheckThroughSessionManagerAndNeverLogsTheSkipWarn(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig.Profile cfg = opencodeCfg("opencode/nemotron-3-ultra-free", null, null);
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, sink);
+        SessionManager sessions = new SessionManager(launcher);
+
+        Logger logger = (Logger) LoggerFactory.getLogger(OpenCodeLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
+            assertNull(acquired.agentSessionId(), "no opencode row yet");
+
+            OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
+                    "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+
+            Optional<MemberSession> after = sessions.get(acquired.paneId());
+            assertEquals("ses_x", after.get().agentSessionId());
+        } finally {
+            logger.detachAppender(appender);
+        }
+
+        assertEquals(1, exhausted.size(),
+                "the mismatch check still runs on the real path with a provisioned worktree: " + exhausted);
+        boolean cannotRunWarn = appender.list.stream()
+                .filter(e -> e.getLevel() == Level.WARN)
+                .anyMatch(e -> e.getFormattedMessage().contains("cannot run"));
+        assertFalse(cannotRunWarn, "a provisioned-worktree spawn must never log the fleetd #267 "
+                + "'cannot run' WARN — the check ran, it was not skipped: " + appender.list);
+    }
+
+    /**
+     * fleetd #267's central defect, reproduced and fixed: {@code OpenCodeLauncher.SessionAwareHandle
+     * .agentSessionId()} is the ONLY caller of {@code checkModelMatch}, and it sits behind the
+     * fleetd #249 worktree gate — so a plain {@code fleet_spawn} with no {@code worktree:true}
+     * (the ticket's "ordinary, expected shape of the large majority of spawns") never reached
+     * {@code checkModelMatch} at all. A test that called {@code checkModelMatch} directly, or built
+     * a {@link OpenCodeLauncher.SessionAwareHandle}/{@link PeerHandle} in isolation, would have
+     * passed on every single day this gap existed — it never drives {@code agentSessionId()}
+     * through the worktree gate the way production does. This test instead drives the REAL
+     * late-resolve path: {@code SessionManager.acquire()} (which calls {@code handle
+     * .agentSessionId()} to build the very first {@code MemberSession}) and a re-poll via {@code
+     * SessionManager.get()} (fleetd #209's retained-handle mechanism) — the exact sequence a live
+     * pane goes through.
+     *
+     * <p>The fix chosen (see {@code OpenCodeLauncher}'s javadoc on the {@code !worktreeProvisioned}
+     * branch) is the WARN path, not a decoupled check: {@code actualModelForSessionId} can only be
+     * keyed safely by a RESOLVED session id (fleetd #234's fix for exactly this false-positive
+     * risk), and without a provisioned worktree no id can ever be safely resolved (fleetd #249) —
+     * re-deriving "whatever is newest in this shared directory" here would silently reintroduce the
+     * false-positive risk #234 fixed. This test proves both halves: a plausible-looking mismatch
+     * row for the shared, non-provisioned cwd never quarantines anything, AND the new one-time,
+     * per-profile WARN replaces the old total silence.
+     */
+    @Test
+    void aSpawnWithoutAProvisionedWorktreeNeverRunsTheModelCheckButWarnsOncePerProfile(
+            @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        // Deliberately NOT provisionedWorkDir(...) / markAsProvisionedWorktree(...): a plain
+        // directory with no .git marker — the exact "fleet_spawn with no worktree:" shape fleetd
+        // #267 is about, and the ordinary shape the ticket says most spawns actually take.
+        Path workDir = Files.createDirectories(configRoot.resolve("shared-cwd"));
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig.Profile cfg = opencodeCfg("opencode/nemotron-3-ultra-free", null, null);
+        List<String> exhausted = new ArrayList<>();
+        ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
+        OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, sink);
+        SessionManager sessions = new SessionManager(launcher);
+
+        Logger logger = (Logger) LoggerFactory.getLogger(OpenCodeLauncher.class);
+        ListAppender<ILoggingEvent> appender = new ListAppender<>();
+        appender.start();
+        logger.addAppender(appender);
+        try {
+            // The real production entrypoint: acquire() calls handle.agentSessionId() itself to
+            // build the very first MemberSession, BEFORE any row exists.
+            MemberSession acquired = sessions.acquire(cfg.profile(), workDir.toString(), null, null);
+            assertNull(acquired.agentSessionId(),
+                    "still refuses to guess an identity for a shared, non-provisioned cwd (fleetd #249)");
+
+            // A row for this exact (shared) directory appears, running a model that WOULD look like
+            // a mismatch against cfg.model() if fleetd trusted the shared-directory heuristic —
+            // exactly the false-positive shape fleetd #234 fixed for the id-resolved case.
+            OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_sibling", workDir.toString(), 1000L,
+                    "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
+
+            // Re-drive the SAME real late-resolve path (fleetd #209) — repeatedly, to also prove
+            // the new WARN fires at most once per profile, not once per poll.
+            Optional<MemberSession> resolved = sessions.get(acquired.paneId());
+            assertNull(resolved.get().agentSessionId(), "still no identity — the gate never opens");
+            sessions.get(acquired.paneId());
+        } finally {
+            logger.detachAppender(appender);
+        }
+
+        assertTrue(exhausted.isEmpty(),
+                "must never quarantine off a shared-directory row it cannot trust as this session's "
+                        + "own — fleetd #234's exact concern, now also for the model check: " + exhausted);
+
+        List<String> skipWarnings = appender.list.stream()
+                .filter(e -> e.getLevel() == Level.WARN)
+                .map(ILoggingEvent::getFormattedMessage)
+                .filter(m -> m.contains("cannot run"))
+                .toList();
+        assertEquals(1, skipWarnings.size(),
+                "exactly one 'cannot run' WARN across acquire() + two get() re-polls — the old code "
+                        + "logged NOTHING here, which is the bug this ticket fixes; got: " + skipWarnings);
+        assertTrue(skipWarnings.get(0).contains(cfg.profile()),
+                "the WARN must name the profile, same treatment discoveryUnavailable already gets: "
+                        + skipWarnings.get(0));
+    }
 }

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

@@ -7,6 +7,7 @@ import java.util.Locale;
 import java.util.Set;
 import java.util.regex.Matcher;
 import java.util.regex.Pattern;
+import java.util.stream.Collectors;
 import org.junit.jupiter.api.DisplayName;
 import org.junit.jupiter.api.Test;
 
@@ -68,6 +69,32 @@ class RestRouteInventoryTest {
             "GET /tasks/{ticket}"
     );
 
+    private static final Pattern ROUTE_CALL =
+            Pattern.compile("app\\.(get|post|delete|put|patch)\\(\\s*\"([^\"]+)\"");
+
+    /**
+     * Drop whole-line comments before scraping. Without this the scrape reads commented-out code as
+     * live: a registration disabled with {@code //} still matched, so the route stayed in the
+     * inventory while the server no longer served it — a silent false PASS, measured on 2026-09-03
+     * by commenting out {@code app.get("/tasks/{ticket}", ...)} and watching this test stay green.
+     * Deleting the same line was caught correctly, so only the commented-out shape was blind.
+     *
+     * <p>Only lines whose first non-blank characters are {@code //}, {@code *} or {@code /*} are
+     * dropped — deliberately NOT every {@code //} anywhere on a line, because that would also cut a
+     * string literal containing {@code //} (a URL) and could silently delete a real registration
+     * sharing that line. The remaining gap is a trailing comment on the same line as real code; no
+     * registration in this file has that shape, and the vacuity test below would catch a scrape that
+     * lost registrations wholesale.
+     */
+    private static String withoutCommentLines(String source) {
+        return source.lines()
+                .filter(line -> {
+                    String t = line.stripLeading();
+                    return !(t.startsWith("//") || t.startsWith("*") || t.startsWith("/*"));
+                })
+                .collect(Collectors.joining("\n"));
+    }
+
     /**
      * Every {@code app.<verb>("<path>")} call in {@link FleetApp}'s source, as {@code "VERB path"}.
      * This matches inside an {@code if (...) { ... }} block just as well as a top-level statement —
@@ -75,8 +102,7 @@ class RestRouteInventoryTest {
      * catches {@code GET /metrics} (registered conditionally on {@code metrics != null}).
      */
     private static Set<String> routesTheServerRegisters() throws Exception {
-        String source = Files.readString(REST_SOURCE);
-        Matcher m = Pattern.compile("app\\.(get|post|delete|put|patch)\\(\\s*\"([^\"]+)\"").matcher(source);
+        Matcher m = ROUTE_CALL.matcher(withoutCommentLines(Files.readString(REST_SOURCE)));
         Set<String> found = new LinkedHashSet<>();
         while (m.find()) {
             found.add(m.group(1).toUpperCase(Locale.ROOT) + " " + m.group(2));

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

@@ -45,6 +45,9 @@ public final class FakeWorktrees implements Worktrees {
     private final List<String> overlayShareOrder = new CopyOnWriteArrayList<>();
     private final Set<String> existingPaths = ConcurrentHashMap.newKeySet();
     private final Set<String> trackedPaths = ConcurrentHashMap.newKeySet();
+    /** Worktree paths that currently exist, mirroring GitWorktrees' {@code Files.exists} check for
+     *  the already-gone case (CB-576 review, fleetd #116). */
+    private final Set<String> worktreePaths = ConcurrentHashMap.newKeySet();
     private final AtomicLong snapshotSeq = new AtomicLong();
     private volatile RuntimeException addFailure;
     private volatile RuntimeException snapshotFailure;
@@ -115,7 +118,15 @@ public final class FakeWorktrees implements Worktrees {
         }
         // The branch already carries a unique nonce, so the derived path is distinct per acquire
         // without an extra counter — keep it a pure function of the branch the test can predict.
-        return prefix + "/" + branch.replace('/', '_');
+        String path = prefix + "/" + branch.replace('/', '_');
+        worktreePaths.add(path);
+        return path;
+    }
+
+    /** Model an operator / {@code git worktree prune} removing the worktree before release. */
+    public FakeWorktrees markGone(String worktreePath) {
+        worktreePaths.remove(worktreePath);
+        return this;
     }
 
     @Override
@@ -125,6 +136,11 @@ public final class FakeWorktrees implements Worktrees {
 
     @Override
     public boolean hasUncommitted(String worktreePath) {
+        // A path that was never added, or was marked gone, is reported clean, mirroring
+        // GitWorktrees' already-gone guard — never an error, so teardown still completes.
+        if (!worktreePaths.contains(worktreePath)) {
+            return false;
+        }
         return dirty;
     }
 

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

@@ -258,6 +258,38 @@ class WorktreeSessionManagerTest {
         }
     }
 
+    /**
+     * CB-576 review (fleetd #116). A worktree that is already gone (operator cleanup,
+     * {@code git worktree prune}, an earlier half-completed release) must not break teardown.
+     * {@code hasUncommitted} reports the missing path clean (mirroring {@code GitWorktrees}), so
+     * {@code release} still runs {@code notifyReleased} (the CB-516 fast-fail for a blocked
+     * {@code fleet_send} caller) and {@code launcher.stop} (so the pane is not orphaned), and falls
+     * through to the already-gone-tolerant {@code remove}. Since CB-581 this all happens because the
+     * notify-and-stop work sits in {@code release}'s {@code finally}/post-try block rather than a
+     * checked branch — this test pins that shape by construction.
+     */
+    @Test
+    void releaseStillStopsPaneAndNotifiesWhenWorktreeIsGone() {
+        FakeHerdr herdr = new FakeHerdr();
+        FakeWorktrees worktrees = new FakeWorktrees().withRepoRoot("/repo").withPrefix("/wt");
+        SessionManager sessions = new SessionManager(workerService(herdr), worktrees);
+        List<SessionManager.ReleaseDetail> released = new java.util.concurrent.CopyOnWriteArrayList<>();
+        sessions.onRelease(released::add);
+        MemberSession s = sessions.acquire("ltms-local", null, "/caller/proj", null,
+                new WorktreeRequest("cb-576g", null));
+
+        worktrees.markGone(s.worktree());
+        sessions.release(s.paneId());
+
+        assertEquals(1, released.size(),
+                "notifyReleased must still fire when the worktree is already gone (CB-516)");
+        assertEquals(s.terminalId(), released.getFirst().terminalId());
+        assertTrue(herdr.called("pane.close"),
+                "the pane must still be stopped when the worktree is already gone");
+        assertEquals(1, worktrees.removeCalls().size(),
+                "release still calls the already-gone-tolerant remove");
+    }
+
     @Test
     void drainAllPreservesWorktreeOfIdleSession() {
         FakeHerdr herdr = new FakeHerdr();