fleetd #249: withhold agentSessionId for a non-provisioned opencode cwd

OpenCodeSessionDiscovery.sessionIdForDirectory keys on the worker's cwd, which
is reliable only when fleetd provisioned a unique git worktree for that
member. Without one (the default no-worktree spawn), the cwd is shared with
other sessions, and "most recently updated row for this directory" can pick a
stranger's session — fleet_list would then hand a lead an agentSessionId that
resumes someone else's conversation.

Move isProvisionedWorktree from ClaudeCodeLauncher to the shared
HerdrPeerLauncher base (both adapters need it now). OpenCodeLauncher.spawn now
refuses a resumeSessionId spawn outright when the target cwd is not a
provisioned worktree (fleetd can never verify or re-report that identity), and
SessionAwareHandle.agentSessionId() withholds the id — returns null rather
than guessing — for any member spawned without one, resumed or not. Corrected
fleet_list/fleet_spawn's tool descriptions, which previously implied
agentSessionId is always a safe resume handle.

CLAUDE.md

@@ -207,6 +207,14 @@ must obey belongs in the charter, not here.
   `fleets-status` (report every fleet that shares one LavinMQ instance).
 - **Never commit** `.mcp.json` (the primary's local copy, flagged `--skip-worktree`) or `wiki/`
   (a submodule with its own remote).
+- **A provisioned worktree neutralizes `.mcp.json`, `opencode.json` and `.autoenv`** — the repo's
+  committed copies would otherwise mount the primary's IDE and forge servers (fleetd #134). The
+  worktree's copy of each is a stub, **not** the repo's real file, so a worker that reads one and
+  reports what it found is reporting on the stub. The daemon logs a per-spawn summary, but the
+  worker cannot see that log. From inside its own worktree a worker — or a lead debugging one —
+  reads the list with `git config --worktree --get-all fleet.neutralizedConfig`, and the
+  consequence with `git config --worktree --get fleet.neutralizedConfigNote`. Never brief a worker
+  to edit one of these files: the edit cannot be committed, and it will not tell you so.
 - **Flows and the error model** — rendezvous, `fleet_ask`, detached delivery, turn-done fallback —
   are diagrammed in `docs/MCP-Contract.md` **§6 only**. The rest of that page is a pre-build design
   doc whose tool names, parameter names and REST paths never caught up with the code, so do not use

fleetd/fleetd.example.yaml

@@ -172,7 +172,11 @@ herdrSocket: ~/.config/herdr/herdr.sock
 #                      skills/MCP/hooks. Omit to leave the worker on the host default.
 #   parityOverlay    → repo-relative paths copied primary→worktree so a worker in a provisioned
 #                      worktree sees the same local config (CB-301-ext). Omit for the default set:
-#                      [.env, .envrc]. (.claude/settings.local.json is NOT in the default — it
+#                      [.env] only (CB-148). .envrc is left out of the default on purpose: it is
+#                      executable shell that direnv runs on every cd, so copying it carries
+#                      behaviour into the worker, not just values, unlike .env. An operator who
+#                      wants it copied can still write parityOverlay: [.env, .envrc] explicitly.
+#                      (.claude/settings.local.json is NOT in the default — it
 #                      pre-approves IDE/tool grants a member must not hold ambiently; CB-525/CB-634.)
 #
 #                      Do NOT add .mcp.json (CB-525). A worker's tools are whatever its launcher
@@ -309,7 +313,7 @@ profiles:
     # errorPattern: "503 Service Unavailable"   # opt-in: classify a backend outage (fleetd #201/#227) — see the key doc above
     # configDir: /Users/me/.ccs/instances/gx10   # CLAUDE_CONFIG_DIR — inherit that profile's skills/MCP
     # cwd: /Users/me/src/myrepo       # pin the working dir; omit to inherit the primary's
-    # parityOverlay: [".env", ".envrc"]   # the default; never add .mcp.json or .claude/settings.local.json — see above
+    # parityOverlay: [".env"]   # the default; add ".envrc" explicitly if you want it copied too (CB-148) — never add .mcp.json or .claude/settings.local.json — see above
     # ideMcpUrl: http://127.0.0.1:29170/index-mcp/streamable-http   # opt-in (CB-634): IDE code intelligence, pinned to the worktree
     # ideProjectDir: fleetd          # CB-634: module dir the IDE opens + the overlay pins (this repo's pom is in fleetd/)
     # ideOpenCommand: env DISPLAY=:10.0 idea {dir}   # CB-634 auto-open: opens {dir} in the IDE at spawn; omit to open by hand

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

@@ -351,7 +351,8 @@ public final class Fleetd {
                 .map(session -> exhaustedPatternsByProfile.get(session.profile()))
                 .orElse(null);
         log.info("backend-exhausted classification (CB-578 stage A): {}",
-                CompletionResolver.coverage(cfg.profiles().keySet(), exhaustedPatternsByProfile.keySet()));
+                CompletionResolver.coverage("exhaustedPattern", cfg.profiles().keySet(),
+                        exhaustedPatternsByProfile.keySet()));
         // fleetd #201 Unit 5: classify a completion-fallback scrape that matches a profile's
         // configured backend-error refusal (a credential outage, a provider 5xx) as a backend error
         // rather than handing it back as a real answer. Compiled once at startup, keyed by profile
@@ -365,13 +366,14 @@ public final class Fleetd {
                 errorPatternsByProfile.put(name, Pattern.compile(profile.errorPattern()));
             }
         });
-        BackendErrorPatternLookup backendErrorPatterns = target -> sessions.roster().stream()
-                .filter(session -> target.equals(session.terminalId()))
-                .findFirst()
-                .map(session -> errorPatternsByProfile.get(session.profile()))
-                .orElse(null);
+        // fleetd #248: extracted to a static factory (see backendErrorPatternLookup below) so a
+        // test can prove main() actually PASSES this into CompletionResolver, not only that the
+        // lookup itself behaves correctly — the exact gap fleetd #248 exists to close.
+        BackendErrorPatternLookup backendErrorPatterns = backendErrorPatternLookup(sessions::roster,
+                errorPatternsByProfile);
         log.info("backend-error classification (fleetd #201 Unit 5): {}",
-                CompletionResolver.coverage(cfg.profiles().keySet(), errorPatternsByProfile.keySet()));
+                CompletionResolver.coverage("errorPattern", cfg.profiles().keySet(),
+                        errorPatternsByProfile.keySet()));
         // CB-578 stage B: on a classification that actually wins, quarantine the exhausted profile's
         // CREDENTIAL — not the profile name — so a profile sharing that credential (e.g. two models
         // on one OpenAI account) is refused too, not just the one that happened to report it. Reads
@@ -424,51 +426,25 @@ public final class Fleetd {
         // construction-order cycle `exhaustionSinkRef` breaks above, broken the same way: a mutable
         // holder set once `pushLoop` exists, read lazily from inside the lambda built here.
         AtomicReference<ReplyPushLoop> pushLoopRef = new AtomicReference<>();
-        // Order: (1) mark the member BACKEND_ERROR; (2) resolve profile/credential through the
-        // roster — fail loud (never Optional.ifPresent, the fleetd #234 lesson applied to this new
-        // sink) and notify the lead via onBackendTargetUnmapped when it cannot be resolved; (3)
-        // record the error in BackendOutagePolicy; (4) on a NEW incident (the record() call that
-        // actually crosses the threshold), tell the lead via onBackendIncident.
-        BackendErrorSink backendErrorSink = (target, matchedLine, reason) -> {
-            sessions.onBackendError(target, reason);
-
-            String profileName = sessions.roster().stream()
-                    .filter(session -> target.equals(session.terminalId()))
-                    .findFirst()
-                    .map(MemberSession::profile)
-                    .orElse(null);
-            FleetConfig.Profile profile = profileName == null ? null : config.get().profiles().get(profileName);
-            if (profile == null) {
-                log.error("backend error on target '{}' ({}) but no profile could be resolved — the "
-                                + "target is not (yet) in the roster — no cool-off applied (fleetd #201 Unit 5)",
-                        target, reason);
-                ReplyPushLoop loop = pushLoopRef.get();
-                if (loop != null) {
-                    loop.onBackendTargetUnmapped(target, reason);
-                }
-                return;
-            }
-            String credentialId = profile.effectiveCredentialId();
-            Optional<BackendOutagePolicy.Incident> incident = outagePolicy.record(credentialId, target, reason);
-            incident.ifPresent(inc -> {
-                List<String> affectedProfiles = config.get().profiles().values().stream()
-                        .filter(p -> credentialId.equals(p.effectiveCredentialId()))
-                        .map(FleetConfig.Profile::profile)
-                        .sorted()
-                        .toList();
-                log.warn("credential '{}' cooling off for {}s after backend errors on {} distinct "
-                                + "target(s) (profile '{}'): {}", credentialId,
-                        inc.remainingCoolOffSeconds(), inc.evidenceCount(), profile.profile(), reason);
-                ReplyPushLoop loop = pushLoopRef.get();
-                if (loop != null) {
-                    loop.onBackendIncident(inc.id(), inc.targets(), credentialId, affectedProfiles,
-                            (int) inc.remainingCoolOffSeconds());
-                }
-            });
-        };
+        // fleetd #248: extracted to a static factory (see backendErrorSink below), public rather
+        // than package-private like the other two factories here, so
+        // dev.ltms.fleet.inject.BackendOutageFlowTest can exercise the REAL production sink
+        // directly instead of a hand-mirrored copy of this lambda — that copy was precisely the
+        // gap fleetd #248 exists to close (see that test's class doc for the history).
+        BackendErrorSink backendErrorSink = backendErrorSink(sessions, () -> config.get().profiles(),
+                outagePolicy, pushLoopRef::get);
         AgentControl agents = router.memberAgents();
+        // Both fleetd#201 Unit 5 (backend-error patterns + sink) and fleetd#241 (the worktree/branch
+        // lookup the fallback report names) land on this one call. The full constructor takes both,
+        // so neither feature is dropped; nowNanos must be passed explicitly to reach it.
+        //
+        // fleetd #248: every argument built specifically for this call (backendErrorPatterns and
+        // backendErrorSink above, and the worktree/branch lookup right here) now comes from a
+        // static factory tested on its own; FleetdCompletionResolverWiringTest source-asserts that
+        // THIS call actually passes them, which is the coverage that was missing before.
         CompletionResolver completion = new CompletionResolver(agents, rendezvous, exhaustedPatterns,
-                exhaustionSink, backendErrorPatterns, backendErrorSink);
+                exhaustionSink, backendErrorPatterns, backendErrorSink, System::nanoTime,
+                worktreeBranchLookup(sessions::roster));
         // CB-113: deliver only to an available worker (its MCP is connected), never its boot window.
         // CB-301: the manager's presence bridge records availability and drives SPAWNING → READY.
         MemberPresence presence = sessions.asPresence();
@@ -762,6 +738,123 @@ public final class Fleetd {
         return target -> presence.isPresent(target) || leads.get().containsKey(target);
     }
 
+    /**
+     * fleetd #248: package-private factory for the member worktree/branch lookup {@link
+     * CompletionResolver} uses to name a fallback report's worktree and branch (fleetd#241).
+     *
+     * <p>Before this ticket the lookup was an anonymous lambda built inline inside {@code main}'s
+     * {@code CompletionResolver} constructor call — provably untested wiring, the whole reason
+     * fleetd #248 exists: dropping that one argument (passing {@code _ -> null} instead) compiled
+     * clean and left every test green. Extracted here, {@code main} now calls this factory instead
+     * of building the lambda inline, and a source assertion on that call site
+     * ({@code FleetdCompletionResolverWiringTest}) proves the argument is still actually passed.
+     *
+     * <p>Takes the roster as a plain {@link Supplier} — not a {@link SessionManager} — so this is
+     * directly testable with a hand-built session list; no real {@code SessionManager} (launcher,
+     * worktrees, …) needs constructing. Follows the same {@code static} factory pattern as
+     * {@link #deliverableTo} above.
+     *
+     * @param roster the live member roster, normally {@code sessions::roster}
+     */
+    static Function<String, CompletionResolver.WorktreeBranch> worktreeBranchLookup(
+            Supplier<List<MemberSession>> roster) {
+        return target -> roster.get().stream()
+                .filter(session -> target.equals(session.terminalId()))
+                .findFirst()
+                .map(session -> new CompletionResolver.WorktreeBranch(session.worktree(), session.branch()))
+                .orElse(null);
+    }
+
+    /**
+     * fleetd #248 / fleetd#201 Unit 5: package-private factory for the per-target backend-error
+     * pattern lookup {@link CompletionResolver} classifies a pane scrape against. Closes over the
+     * live roster (to resolve a target to a profile) and {@code errorPatternsByProfile} (each
+     * profile's configured {@code errorPattern}, already compiled by the caller — the same map also
+     * feeds the coverage log next to where this is called) — nothing else, so it is directly
+     * testable. See {@link #worktreeBranchLookup} above for why this ticket exists and why the
+     * factory takes a roster {@link Supplier} rather than a {@link SessionManager}.
+     *
+     * @param roster                 the live member roster, normally {@code sessions::roster}
+     * @param errorPatternsByProfile every profile that has an {@code errorPattern} configured,
+     *                               keyed by profile name
+     */
+    static BackendErrorPatternLookup backendErrorPatternLookup(Supplier<List<MemberSession>> roster,
+            Map<String, Pattern> errorPatternsByProfile) {
+        return target -> roster.get().stream()
+                .filter(session -> target.equals(session.terminalId()))
+                .findFirst()
+                .map(session -> errorPatternsByProfile.get(session.profile()))
+                .orElse(null);
+    }
+
+    /**
+     * fleetd #248 / fleetd#201 Unit 5: factory for the production {@link BackendErrorSink} — the
+     * collaborator {@link CompletionResolver} notifies when a pane-scrape classification actually
+     * resolves a waiter as a backend error. Order: (1) mark the member BACKEND_ERROR; (2) resolve
+     * profile/credential through the roster — fail loud (never {@code Optional.ifPresent}, the
+     * fleetd #234 lesson applied to this sink) and notify the lead via {@code
+     * onBackendTargetUnmapped} when it cannot be resolved; (3) record the error in {@code
+     * outagePolicy}; (4) on a NEW incident (the {@code record()} call that actually crosses the
+     * threshold), tell the lead via {@code onBackendIncident}.
+     *
+     * <p>{@code public}, unlike {@link #worktreeBranchLookup} and {@link #backendErrorPatternLookup}
+     * above: {@code dev.ltms.fleet.inject.BackendOutageFlowTest} exercises this exact object as the
+     * real, wired production path, replacing what its own class doc used to call out as a
+     * hand-mirrored copy of this lambda ("mirrors {@code Fleetd.main}'s {@code backendErrorSink}
+     * lambda line-for-line") — that copy proved only itself, never that {@code main} still wires the
+     * real thing. That was precisely the gap fleetd #248 exists to close.
+     *
+     * @param sessions the session registry; both read (roster) and written (onBackendError)
+     * @param profiles the live profile map, normally {@code () -> config.get().profiles()} in
+     *                 {@code main}, or a fixed test map via {@code () -> profiles}
+     * @param pushLoop the lead-nudge loop, read lazily: {@code main} builds this sink before the
+     *                 real {@link ReplyPushLoop} exists (a genuine construction-order cycle, broken
+     *                 the same way {@code exhaustionSinkRef} is a few lines above it), so a
+     *                 {@link Supplier} reads whatever {@code main} has filled in by the time a real
+     *                 backend error fires
+     */
+    public static BackendErrorSink backendErrorSink(SessionManager sessions,
+            Supplier<Map<String, FleetConfig.Profile>> profiles, BackendOutagePolicy outagePolicy,
+            Supplier<ReplyPushLoop> pushLoop) {
+        return (target, matchedLine, reason) -> {
+            sessions.onBackendError(target, reason);
+
+            String profileName = sessions.roster().stream()
+                    .filter(session -> target.equals(session.terminalId()))
+                    .findFirst()
+                    .map(MemberSession::profile)
+                    .orElse(null);
+            FleetConfig.Profile profile = profileName == null ? null : profiles.get().get(profileName);
+            if (profile == null) {
+                log.error("backend error on target '{}' ({}) but no profile could be resolved — the "
+                                + "target is not (yet) in the roster — no cool-off applied (fleetd #201 Unit 5)",
+                        target, reason);
+                ReplyPushLoop loop = pushLoop.get();
+                if (loop != null) {
+                    loop.onBackendTargetUnmapped(target, reason);
+                }
+                return;
+            }
+            String credentialId = profile.effectiveCredentialId();
+            Optional<BackendOutagePolicy.Incident> incident = outagePolicy.record(credentialId, target, reason);
+            incident.ifPresent(inc -> {
+                List<String> affectedProfiles = profiles.get().values().stream()
+                        .filter(p -> credentialId.equals(p.effectiveCredentialId()))
+                        .map(FleetConfig.Profile::profile)
+                        .sorted()
+                        .toList();
+                log.warn("credential '{}' cooling off for {}s after backend errors on {} distinct "
+                                + "target(s) (profile '{}'): {}", credentialId,
+                        inc.remainingCoolOffSeconds(), inc.evidenceCount(), profile.profile(), reason);
+                ReplyPushLoop loop = pushLoop.get();
+                if (loop != null) {
+                    loop.onBackendIncident(inc.id(), inc.targets(), credentialId, affectedProfiles,
+                            (int) inc.remainingCoolOffSeconds());
+                }
+            });
+        };
+    }
+
     /** Injection seam for {@link #selectReplyInbox}: production binds {@link AmqpReplyInbox#open}. */
     @FunctionalInterface
     interface AmqpOpener {

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

@@ -259,7 +259,13 @@ public record FleetConfig(
      * @param cwd            fixed working directory for this profile's workers (CB-112 "told otherwise");
      *                       {@code null}/blank → inherit the primary's cwd, else the daemon's
      * @param parityOverlay  repo-relative paths copied primary→worktree for config parity; null/empty
-     *                       defaults to a sensible set of local config files.
+     *                       defaults to {@code [.env]} only (CB-148 point 2). {@code .env} is data — a
+     *                       copy of it can only carry values. {@code .envrc} is executable shell that
+     *                       {@code direnv} evaluates on every {@code cd}, so copying it moves behaviour
+     *                       into the worker, not just values, and that is a different risk from copying
+     *                       data. It is deliberately left out of the default for that reason. The knob
+     *                       still supports it: an operator who wants it copied writes
+     *                       {@code parityOverlay: [.env, .envrc]} explicitly and owns that choice.
      *                       <p><strong>Every overlay path must be gitignored or tracked-and-skipped.</strong>
      *                       CB-576 made {@code release()} preserve a worktree that {@code git status
      *                       --porcelain} reports as dirty, and it deliberately counts untracked files —
@@ -269,11 +275,10 @@ public record FleetConfig(
      *                       worktrees then accumulate with no error anywhere.
      *                       <p>Checked on 2026-08-15 (CB-581): inert as configured. Tracked overlay
      *                       files carry {@code --skip-worktree} so {@code --porcelain} cannot see them,
-     *                       {@code fleetd.yaml} is gitignored, and the default pair {@code .env} /
-     *                       {@code .envrc} does not exist in this repo. Note the default applies to
-     *                       <em>every</em> profile, so creating either file at the repo root is enough
-     *                       to make it live. Add a new overlay path to {@code .gitignore} in the same
-     *                       change that adds it here.
+     *                       {@code fleetd.yaml} is gitignored, and {@code .env} does not exist in this
+     *                       repo. Note the default applies to <em>every</em> profile, so creating
+     *                       {@code .env} at the repo root is enough to make it live. Add a new overlay
+     *                       path to {@code .gitignore} in the same change that adds it here.
      *                       <p>CB-578 stage C raised the stakes: a preserved dirty worktree is now also
      *                       committed to {@code refs/wip/<branch>} via {@code git add -A}. The overlay
      *                       carries the primary's own environment files into the worktree, so a
@@ -421,7 +426,7 @@ public record FleetConfig(
             // servers do not exist there to be enabled. This is defence in depth, not a live fix: the
             // worker stays isolated only because this separate mechanism already removes the servers.
             parityOverlay = (parityOverlay == null || parityOverlay.isEmpty())
-                    ? List.of(".env", ".envrc")
+                    ? List.of(".env")
                     : List.copyOf(parityOverlay);
             // gitTokenEnv stays null when unset (opt-in). gitHostEnv defaults so operators enabling
             // checkpoints need only set gitTokenEnv; it is injected only alongside a resolved token.

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

@@ -14,6 +14,7 @@ import java.util.TreeSet;
 import java.util.concurrent.CompletableFuture;
 import java.util.concurrent.ConcurrentHashMap;
 import java.util.function.LongSupplier;
+import java.util.function.Function;
 import java.util.regex.Pattern;
 
 /**
@@ -85,6 +86,17 @@ public final class CompletionResolver implements TurnListener {
     private static final String CLIPPED_PANE_TAIL_MARKER =
             "[Pane tail clipped: member did not call fleet_reply.]";
 
+    /** A pane echo must be this large before it can replace a completion report. */
+    static final int ECHO_MIN_CHARS = 400;
+
+    /** Normalised TUI chrome may add this many characters to an otherwise echoed brief. */
+    static final int MAX_ECHO_EXCESS_CHARS = 160;
+
+    /** The explicit result returned instead of a lead's echoed injected brief. */
+    public static final String NO_REPORT_PREFIX = "[no report — the member ended its turn without fleet_reply, "
+            + "and the pane still shows the injected brief. Nothing was produced on the pane. Check the "
+            + "member's worktree and branch for committed work before re-delegating.";
+
     private final AgentControl agents;
     private final Rendezvous rendezvous;
     private final ExhaustedPatternLookup exhaustedPatterns;
@@ -92,6 +104,11 @@ public final class CompletionResolver implements TurnListener {
     private final BackendErrorPatternLookup backendErrorPatterns;
     private final BackendErrorSink backendErrorSink;
     private final LongSupplier nowNanos;
+    private final Function<String, WorktreeBranch> worktreeBranches;
+
+    /** Known member location, used only to guide a lead after an echoed brief. */
+    public record WorktreeBranch(String worktree, String branch) {
+    }
 
     /**
      * Per-target record of the turn currently in flight: the exact {@link Rendezvous} waiter its
@@ -108,7 +125,8 @@ public final class CompletionResolver implements TurnListener {
      * the other half of the {@link #MIN_TURN_NANOS} floor check, compared against a fresh reading at
      * resolution time.
      */
-    record InFlight(CompletableFuture<Rendezvous.Resolution> waiter, String baseline, long deliveredAtNanos) {
+    record InFlight(CompletableFuture<Rendezvous.Resolution> waiter, String baseline, long deliveredAtNanos,
+                    String injectedText) {
 
         /**
          * Convenience for tests exercising scrape/suppression logic that don't care about turn
@@ -116,7 +134,11 @@ public final class CompletionResolver implements TurnListener {
          * Not used by production code — {@link #captureBaseline} always records a real reading.
          */
         InFlight(CompletableFuture<Rendezvous.Resolution> waiter, String baseline) {
-            this(waiter, baseline, Long.MIN_VALUE / 2);
+            this(waiter, baseline, Long.MIN_VALUE / 2, null);
+        }
+
+        InFlight(CompletableFuture<Rendezvous.Resolution> waiter, String baseline, long deliveredAtNanos) {
+            this(waiter, baseline, deliveredAtNanos, null);
         }
     }
 
@@ -139,11 +161,18 @@ public final class CompletionResolver implements TurnListener {
      * (fleetd#201 Unit 5).
      */
     public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
-                              ExhaustionSink exhaustionSink) {
+                               ExhaustionSink exhaustionSink) {
         this(agents, rendezvous, exhaustedPatterns, exhaustionSink,
                 BackendErrorPatternLookup.legacy(), BackendErrorSink.none(), System::nanoTime);
     }
 
+    /** Production constructor with a lookup for the member worktree and branch. */
+    public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
+                              ExhaustionSink exhaustionSink, Function<String, WorktreeBranch> worktreeBranches) {
+        this(agents, rendezvous, exhaustedPatterns, exhaustionSink,
+                BackendErrorPatternLookup.legacy(), BackendErrorSink.none(), System::nanoTime, worktreeBranches);
+    }
+
     /**
      * Transition constructor (fleetd#201 Unit 1): same legacy backend-error defaults as the 4-arg
      * constructor above, but with the injectable clock. Kept so existing fleetd#164 timing tests
@@ -172,7 +201,7 @@ public final class CompletionResolver implements TurnListener {
                        ExhaustionSink exhaustionSink, BackendErrorPatternLookup backendErrorPatterns,
                        BackendErrorSink backendErrorSink) {
         this(agents, rendezvous, exhaustedPatterns, exhaustionSink, backendErrorPatterns, backendErrorSink,
-                System::nanoTime);
+                System::nanoTime, _ -> null);
     }
 
     /**
@@ -185,8 +214,17 @@ public final class CompletionResolver implements TurnListener {
      * package.
      */
     public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
-                       ExhaustionSink exhaustionSink, BackendErrorPatternLookup backendErrorPatterns,
-                       BackendErrorSink backendErrorSink, LongSupplier nowNanos) {
+                        ExhaustionSink exhaustionSink, BackendErrorPatternLookup backendErrorPatterns,
+                        BackendErrorSink backendErrorSink, LongSupplier nowNanos) {
+        this(agents, rendezvous, exhaustedPatterns, exhaustionSink, backendErrorPatterns, backendErrorSink,
+                nowNanos, _ -> null);
+    }
+
+    /** Full constructor with injectable clock and member location lookup. */
+    public CompletionResolver(AgentControl agents, Rendezvous rendezvous, ExhaustedPatternLookup exhaustedPatterns,
+                        ExhaustionSink exhaustionSink, BackendErrorPatternLookup backendErrorPatterns,
+                        BackendErrorSink backendErrorSink, LongSupplier nowNanos,
+                        Function<String, WorktreeBranch> worktreeBranches) {
         this.agents = agents;
         this.rendezvous = rendezvous;
         this.exhaustedPatterns = Objects.requireNonNull(exhaustedPatterns, "exhaustedPatterns");
@@ -194,6 +232,7 @@ public final class CompletionResolver implements TurnListener {
         this.backendErrorPatterns = Objects.requireNonNull(backendErrorPatterns, "backendErrorPatterns");
         this.backendErrorSink = Objects.requireNonNull(backendErrorSink, "backendErrorSink");
         this.nowNanos = Objects.requireNonNull(nowNanos, "nowNanos");
+        this.worktreeBranches = Objects.requireNonNull(worktreeBranches, "worktreeBranches");
     }
 
     @Override
@@ -223,7 +262,7 @@ public final class CompletionResolver implements TurnListener {
             baseline = null; // fail open: no baseline ⇒ no suppression
             log.debug("delivery baseline for {} failed: {}", target, e.getMessage());
         }
-        inFlight.put(target, new InFlight(waiter, baseline, nowNanos.getAsLong()));
+        inFlight.put(target, new InFlight(waiter, baseline, nowNanos.getAsLong(), token.injectedText()));
     }
 
     /** The turn currently baselined for {@code target}, or {@code null} — a test hook for the captureBaseline path. */
@@ -369,6 +408,9 @@ public final class CompletionResolver implements TurnListener {
             return;
         }
         String completion = clipped ? tail + "\n" + CLIPPED_PANE_TAIL_MARKER : tail;
+        if (echoesInjectedBrief(tail, turn.injectedText())) {
+            completion = noReportMessage(target) + (clipped ? "\n" + CLIPPED_PANE_TAIL_MARKER : "");
+        }
         if (rendezvous.resolveCompletion(waiter, completion)) {
             inFlight.remove(target, turn);
             if (clipped) {
@@ -381,6 +423,40 @@ public final class CompletionResolver implements TurnListener {
         }
     }
 
+    /**
+     * A full echoed brief is at least 400 normalised characters. A scrape that contains the brief may
+     * add no more than 160 normalised characters of TUI chrome. This accepts harmless status text, but
+     * preserves a real report that restates the full brief before adding substantive content.
+     */
+    static boolean echoesInjectedBrief(String scrape, String injectedText) {
+        String normalScrape = normalize(scrape);
+        String normalInjected = normalize(injectedText);
+        if (normalScrape.length() < ECHO_MIN_CHARS || normalInjected.length() < ECHO_MIN_CHARS) {
+            return false;
+        }
+        if (normalInjected.contains(normalScrape)) {
+            return true;
+        }
+        return normalScrape.contains(normalInjected)
+                && normalScrape.length() - normalInjected.length() <= MAX_ECHO_EXCESS_CHARS;
+    }
+
+    private static String normalize(String text) {
+        return text == null ? "" : text.toLowerCase().replaceAll("[^a-z0-9]+", "");
+    }
+
+    private String noReportMessage(String target) {
+        WorktreeBranch location = worktreeBranches.apply(target);
+        if (location == null || (location.worktree() == null && location.branch() == null)) {
+            return NO_REPORT_PREFIX + "]";
+        }
+        String locationText = location.worktree() == null ? "" : " worktree=" + location.worktree();
+        if (location.branch() != null) {
+            locationText += " branch=" + location.branch();
+        }
+        return NO_REPORT_PREFIX + locationText + "]";
+    }
+
     /**
      * fleetd#211: the raw-scrape fallback classification, run only when {@link #lastAssistantBlock}
      * found nothing usable (see the call site in {@link #resolve}). Mirrors the two classifications
@@ -544,9 +620,9 @@ public final class CompletionResolver implements TurnListener {
      * @param allProfiles        every configured profile name
      * @param configuredProfiles the subset of {@code allProfiles} that carry an exhausted pattern
      */
-    public static String coverage(Set<String> allProfiles, Set<String> configuredProfiles) {
+    public static String coverage(String patternKey, Set<String> allProfiles, Set<String> configuredProfiles) {
         if (configuredProfiles.isEmpty()) {
-            return "off (no profile has an exhaustedPattern configured; profiles: " + sorted(allProfiles) + ")";
+            return "off (no profile has an " + patternKey + " configured; profiles: " + sorted(allProfiles) + ")";
         }
         Set<String> unconfigured = new TreeSet<>(allProfiles);
         unconfigured.removeAll(configuredProfiles);

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

@@ -10,6 +10,7 @@ import dev.ltms.fleet.herdr.Agent;
 import dev.ltms.fleet.metrics.FleetMetrics;
 import dev.ltms.fleet.metrics.Metrics;
 import dev.ltms.fleet.inject.MemberPresence;
+import dev.ltms.fleet.inject.CompletionResolver;
 import dev.ltms.fleet.herdr.HerdrException;
 import dev.ltms.fleet.msg.LeadChannel;
 import dev.ltms.fleet.msg.LeadMessage;
@@ -578,8 +579,9 @@ public final class FleetMcp {
             case REPLIED -> text(r.text());
             // The worker's turn finished but it never called fleet_reply — hand back the scraped
             // transcript tail, flagged so the primary knows it isn't a structured reply.
-            case COMPLETED_UNREPLIED -> text(
-                    "[worker finished without a structured fleet_reply — transcript tail follows]\n" + r.text());
+            case COMPLETED_UNREPLIED -> text(r.text().startsWith(CompletionResolver.NO_REPORT_PREFIX)
+                    ? r.text()
+                    : "[worker finished without a structured fleet_reply — transcript tail follows]\n" + r.text());
             // The worker ran the turn then wedged (CB-109) — surface the error context.
             case WORKER_FAILED -> text("[worker failed — turn ended in an unrecoverable state]\n" + r.text());
             // The backend refused on a subscription usage limit (CB-578 stage A) — the worker's
@@ -694,6 +696,7 @@ public final class FleetMcp {
         }
         return switch (v.phase()) {
             case DONE -> text(v.replySource() != null && v.replySource().equals("transcript")
+                            && !v.reply().startsWith(CompletionResolver.NO_REPORT_PREFIX)
                     ? "[done — worker finished without a structured fleet_reply; transcript tail follows]\n" + v.reply()
                     : v.reply());
             case PENDING -> text("[pending — " + v.detail() + "]");
@@ -1250,7 +1253,11 @@ public final class FleetMcp {
                         + "worktree:<ticket-slug> to provision an isolated git worktree. Pass resumeSessionId "
                         + "to relaunch onto a prior conversation instead of starting cold — this requires an "
                         + "explicit profile whose backend supports it (fleet_list shows agentSessionId for "
-                        + "resumable members), and is refused otherwise rather than silently starting fresh. "
+                        + "resumable members; it is absent for a member fleetd cannot reliably re-identify, "
+                        + "e.g. an opencode member spawned without a worktree), and is refused otherwise "
+                        + "rather than silently starting fresh. For an opencode profile, resumeSessionId "
+                        + "itself also requires worktree:true/<slug> on THIS spawn — without one fleetd can "
+                        + "never re-verify which conversation it actually resumed (fleetd #249). "
                         + "sessionName gives the member a display name in its own UI when the backend supports "
                         + "one. Returns the member's sessionId (use with fleet_send) and paneId (use with "
                         + "fleet_stop).",
@@ -1261,7 +1268,7 @@ public final class FleetMcp {
                         "worktree", Map.of("type", "string", "description", "'true' or a ticket slug — requests an isolated git worktree"),
                         "ticket", stringProp("Ticket slug when worktree:true"),
                         "sessionName", stringProp("Logical display name for the member's own session, when its backend supports one"),
-                        "resumeSessionId", stringProp("A prior member's agentSessionId (from fleet_list) to resume — requires an explicit profile that supports it")),
+                        "resumeSessionId", stringProp("A prior member's agentSessionId (from fleet_list) to resume — requires an explicit profile that supports it, and (for opencode) a worktree on this spawn too")),
                         List.of()));
     }
 
@@ -1282,9 +1289,14 @@ public final class FleetMcp {
                         + "discover a peer lead without being told its address. 'members' are the "
                         + "sessions delegated to — each with sessionId, paneId, role (architect/dev/"
                         + "reviewer), profile (the backend it runs on), state, optional "
-                        + "worktree/branch/owner/agentSessionId (the id to pass as fleet_spawn's "
-                        + "resumeSessionId to relaunch onto that same conversation, when the backend "
-                        + "supports it), and live herdr status. An empty 'members' "
+                        + "worktree/branch/owner/agentSessionId, and live herdr status. agentSessionId, "
+                        + "when present, is the id to pass as fleet_spawn's resumeSessionId to relaunch "
+                        + "onto that same conversation. It is ABSENT — not a guess — for a member fleetd "
+                        + "cannot reliably re-identify: some backends (e.g. opencode) resolve it from the "
+                        + "member's working directory, which only uniquely identifies a member when it "
+                        + "was spawned into its own fleetd-provisioned worktree (worktree:true/<slug>); a "
+                        + "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 "

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

@@ -1,5 +1,8 @@
 package dev.ltms.fleet.member;
 
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.fasterxml.jackson.databind.node.ObjectNode;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.guard.SubscriptionGuard;
 import dev.ltms.fleet.herdr.Agent;
@@ -14,6 +17,8 @@ import java.io.IOException;
 import java.io.UncheckedIOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
+import java.nio.file.StandardCopyOption;
+import java.nio.file.attribute.PosixFileAttributeView;
 import java.util.EnumSet;
 import java.util.List;
 import java.util.Map;
@@ -47,6 +52,21 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
 
     private static final Logger log = LoggerFactory.getLogger(ClaudeCodeLauncher.class);
 
+    /** JSON codec for the additive workspace-trust seed (fleetd #149) — Jackson's default settings. */
+    private static final ObjectMapper TRUST_JSON = new ObjectMapper();
+
+    /**
+     * Serialises every {@link #seedTrustDialog} read-modify-write for the whole daemon process.
+     * Two claude-code spawns starting at once are normal (fleetd runs several members in parallel
+     * routinely) and both would otherwise read the same {@code .claude.json}, add their own entry
+     * to their own in-memory copy, and write — the second write wins and the first spawn's trust
+     * entry silently disappears. A single process-wide lock is enough because every spawn on this
+     * daemon runs in this one JVM; it does not protect against a second daemon process or the
+     * operator's own Claude Code process writing at the same instant, which {@link #writeAtomically}
+     * covers instead (each writer only ever sees a fully-old or fully-new file, never a torn one).
+     */
+    private static final Object TRUST_JSON_LOCK = new Object();
+
     private final SubscriptionGuard guard;
 
     /**
@@ -263,6 +283,10 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
         putIfPresent(workerEnv, "ANTHROPIC_MODEL", cfg.model());
         putIfPresent(workerEnv, "CLAUDE_CONFIG_DIR", cfg.configDir());
         applyGitToken(workerEnv, cfg);
+        // fleetd #149: seed the workspace-trust entry BEFORE this spawn ever reaches herdr — see
+        // seedTrustDialog for why, and isProvisionedWorktree for why this is gated to a worktree
+        // fleetd itself provisioned (never a real checkout, never an un-configured fallback cwd).
+        seedTrustDialog(cfg.configDir(), spec.cwd());
 
         // CB-547a: Claude Code can MINT its own session id, so fleetd chooses it — a fresh spawn
         // gets a UUID we pass as --session-id and return from agentSessionId(), so the resume
@@ -412,12 +436,12 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
      */
     private static void writeIdeOverlay(String cwd, String projectPath) {
         try {
-            Path dotGit = Path.of(cwd, ".git");
-            if (!Files.isRegularFile(dotGit)) {
+            if (!isProvisionedWorktree(cwd)) {
                 // Not a provisioned worktree (primary's real checkout has a .git directory, or the
                 // cwd is not a repo at all). Never write into it.
                 return;
             }
+            Path dotGit = Path.of(cwd, ".git");
             // The overlay FILE lives at the worktree root (claude-code's cwd), but its CONTENT pins
             // project_path to the module dir the IDE opened (projectPath), not the worktree root.
             Files.writeString(Path.of(cwd, "CLAUDE.local.md"), PeerLauncher.ideOverlayText(projectPath));
@@ -447,6 +471,163 @@ public final class ClaudeCodeLauncher extends HerdrPeerLauncher {
         }
     }
 
+    /**
+     * fleetd #149: pre-seed the workspace-trust entry for {@code cwd} in Claude Code's own config
+     * file, BEFORE this launch ever reaches herdr (called from {@link #buildLaunch}, which always
+     * runs before the base class starts the process). Claude Code asks an interactive, un-timed
+     * "Is this a project you created or one you trust?" the first time it starts in a directory it
+     * has not seen before, and every {@code worktree: true} spawn lands in a brand-new directory —
+     * so without this seed the member sits on that dialog forever, never mounts the bridge MCP, and
+     * never calls {@code fleet_reply}. herdr reports it as healthy the whole time ({@code
+     * agent_status: blocked}, {@code interactive_ready: true}), so nothing else catches it. Measured
+     * live on fleet01 2026-08-23: an unseeded fresh cwd sat on the dialog indefinitely; a seeded one
+     * reached {@code idle} clean.
+     *
+     * <p>This is not a new grant — the operator already trusted this repo by configuring the
+     * profile against it, and a worktree is a checkout of that same repo.
+     *
+     * <p>The key Claude Code reads is per-project, in {@code .claude.json}: {@code
+     * projects.<cwd>.hasTrustDialogAccepted}. The file lives at {@code <configDir>/.claude.json}
+     * when the profile sets {@code CLAUDE_CONFIG_DIR} (mirrors this launcher's own env var above),
+     * else the default {@code ~/.claude.json} — the same file Claude Code itself would read either
+     * way, so this seeds exactly what the spawned peer is about to open.
+     *
+     * <p><b>Additive, not a rewrite.</b> {@code .claude.json} is large (tens of KB, dozens of
+     * projects) and Claude Code itself rewrites it while running, so this reads the file as a JSON
+     * tree (missing or unreadable → treated as an empty object) and changes only
+     * {@code projects.<cwd>.hasTrustDialogAccepted} / {@code .hasCompletedProjectOnboarding} —
+     * every other top-level key and every other project entry is written back untouched. Only the
+     * one project entry for {@code cwd} is replaced/created; an existing entry for a DIFFERENT cwd
+     * (or the operator's own project history) is never touched.
+     *
+     * <p>Best-effort, like {@link #writeIdeOverlay}: a failure here (unwritable configDir, a
+     * corrupt existing file, …) must never fail the spawn — it is logged at debug and swallowed. A
+     * peer that starts without the seed still starts; it just may hit the dialog fleetd #149
+     * describes.
+     *
+     * <p><b>Gated to a provisioned worktree</b> ({@link HerdrPeerLauncher#isProvisionedWorktree}) — see that
+     * method's javadoc for the incident that made this gate mandatory, not optional: this must
+     * never run against a real checkout or an un-configured fallback cwd, only the exact
+     * always-fresh-directory population fleetd #149 describes.
+     *
+     * <p><b>Atomic and lock-protected.</b> {@code .claude.json} is a live file — Claude Code itself
+     * rewrites it while running, and this daemon routinely spawns several members at once, each
+     * calling this method for its own cwd. Every write goes through {@link #writeAtomically} (a
+     * sibling-temp-file + {@code ATOMIC_MOVE}, never a truncate-in-place) so a crash mid-write or a
+     * concurrent reader never observes a half-written file, and through {@link #TRUST_JSON_LOCK} so
+     * two concurrent spawns' entries both survive instead of the second write silently discarding
+     * the first. Both exist because of a real incident: see {@link HerdrPeerLauncher#isProvisionedWorktree}'s javadoc
+     * and {@link #writeAtomically}'s javadoc.
+     *
+     * @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
+     *                  up for itself once it starts there
+     */
+    private static void seedTrustDialog(String configDir, String cwd) {
+        if (!isProvisionedWorktree(cwd)) {
+            return;
+        }
+        Path target = (configDir == null || configDir.isBlank())
+                ? Path.of(System.getProperty("user.home"), ".claude.json")
+                : Path.of(configDir, ".claude.json");
+        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;
+                    }
+                }
+                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);
+                }
+                project.put("hasTrustDialogAccepted", true);
+                project.put("hasCompletedProjectOnboarding", true);
+                writeAtomically(target, TRUST_JSON.writerWithDefaultPrettyPrinter().writeValueAsString(root));
+            } catch (Exception e) {
+                log.debug("cannot seed workspace-trust entry for cwd '{}' into '{}'", cwd, target, e);
+            }
+        }
+    }
+
+    /**
+     * 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
+     * one filesystem — a different directory could mean a different filesystem), then
+     * {@link StandardCopyOption#ATOMIC_MOVE} it into place. A reader — Claude Code itself, or
+     * another {@code seedTrustDialog} call — only ever observes the fully-old file or the
+     * fully-new one, never a truncated or half-written one.
+     *
+     * <p><b>fleetd #149 incident.</b> The original implementation used
+     * {@code Files.writeString(target, content)} directly, which truncates {@code target} in place
+     * before writing the replacement bytes. Combined with an ungated {@code cwd} (see
+     * {@link HerdrPeerLauncher#isProvisionedWorktree}'s javadoc), a mutation-testing run hit that truncation window
+     * against the operator's real {@code ~/.claude.json} and left it at 178 bytes. The gate closes
+     * <em>which file</em> this can ever target; this closes <em>how</em> the target is written, so
+     * that even a legitimate write against a real, live, concurrently-read {@code .claude.json}
+     * cannot leave it observably empty or partial.
+     *
+     * <p>Preserves {@code target}'s existing POSIX permissions (Claude Code ships {@code
+     * .claude.json} as {@code 0600}) when the filesystem reports them; a freshly created temp file
+     * already defaults to owner-only permissions on a POSIX filesystem, so a first-ever write (no
+     * existing {@code target}) is no less private without this. On a non-POSIX filesystem (e.g.
+     * Windows) the permission copy is a silent no-op rather than a failure.
+     *
+     * <p>Package-visible (not {@code private}) so a test can drive it directly with a concurrent
+     * reader thread and prove the torn-file property this method exists for — the LOCK in
+     * {@link #seedTrustDialog} already fully serialises every call this launcher itself makes, so a
+     * test that only ever goes through {@code seedTrustDialog}/{@code spawn()} could never observe
+     * a torn file regardless of whether this method is atomic; it would be proving the lock, not
+     * this method. Atomicity's actual job is protecting against a writer the lock cannot reach at
+     * all — a second daemon process, or the operator's own live Claude Code — so the test for it
+     * has to reach this method on its own.
+     */
+    static void writeAtomically(Path target, String content) throws IOException {
+        Path parent = target.getParent();
+        Path tmp = Files.createTempFile(parent, target.getFileName() + ".", ".tmp");
+        try {
+            Files.writeString(tmp, content);
+            copyPosixPermissionsIfPresent(target, tmp);
+            Files.move(tmp, target, StandardCopyOption.ATOMIC_MOVE, StandardCopyOption.REPLACE_EXISTING);
+        } catch (IOException e) {
+            Files.deleteIfExists(tmp);
+            throw e;
+        }
+    }
+
+    /** Copy {@code target}'s POSIX permissions onto {@code tmp}, or no-op where either is unsupported. */
+    private static void copyPosixPermissionsIfPresent(Path target, Path tmp) {
+        try {
+            if (!Files.isRegularFile(target)) {
+                return; // nothing to inherit from — first-ever write, temp file's own default stands
+            }
+            PosixFileAttributeView view = Files.getFileAttributeView(target, PosixFileAttributeView.class);
+            if (view == null) {
+                return; // non-POSIX filesystem — nothing this JVM can read/set here
+            }
+            Files.setPosixFilePermissions(tmp, Files.getPosixFilePermissions(target));
+        } catch (IOException e) {
+            log.debug("cannot preserve permissions of '{}' onto its replacement", target, e);
+        }
+    }
+
     /** {@code s}, or {@code null} when {@code s} is null/blank — the charter-presence test used above. */
     private static String nonBlank(String s) {
         return (s == null || s.isBlank()) ? null : s;

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

@@ -358,6 +358,42 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
         return Files.isRegularFile(candidate) ? candidate : null;
     }
 
+    /**
+     * Whether {@code cwd} is a fleetd-provisioned git worktree — signalled the same way
+     * {@code ClaudeCodeLauncher#writeIdeOverlay} already gates on: a {@code .git} that is a
+     * <strong>regular file</strong> holding a {@code gitdir:} pointer, as opposed to a real
+     * checkout's {@code .git} <strong>directory</strong>. {@code null}/blank never qualifies.
+     *
+     * <p>Shared by every write (and, since fleetd #249, every identity read) that must land only
+     * in a worktree fleetd itself created for a member — never in a real checkout, an arbitrary
+     * configured directory, or (see the incident below) the daemon's own fallback cwd. Package-
+     * private (not {@code protected}) on purpose: {@link ClaudeCodeLauncher} and
+     * {@link OpenCodeLauncher} both call it, and same-package visibility is enough — no subclass
+     * outside this package needs it.
+     *
+     * <p><b>fleetd #149 incident.</b> {@code ClaudeCodeLauncher#seedTrustDialog} originally ran
+     * unconditionally on any non-blank {@code cwd}. Most of that launcher's OWN tests spawn a
+     * profile with no {@code cwd} configured, so the base class's {@code resolveCwd} falls
+     * through to the real {@code user.dir} — and with no {@code configDir} either (also the
+     * common case in that file's fixtures), the seed's target falls through the same way to the
+     * real {@code ~/.claude.json}. Running this repo's own test suite corrupted the operator's
+     * actual config file (it shrank from ~72 KB to a single seeded entry) the first time a
+     * mutation happened to make the write non-additive. Gating both cwd-targeted writes on "this
+     * is a worktree fleetd provisioned" — exactly the population fleetd #149 describes
+     * ({@code worktree: true} always lands in a brand-new directory) — makes that class of write
+     * impossible against a real checkout or an untouched fallback cwd, in production or in tests.
+     *
+     * <p><b>fleetd #249.</b> The same reasoning extends to a READ: {@code
+     * OpenCodeSessionDiscovery#sessionIdForDirectory} keys on {@code directory}, a heuristic that
+     * is only reliable when the directory is unique to this member — i.e., exactly the population
+     * this gate identifies. {@link OpenCodeLauncher} uses it to withhold {@code agentSessionId()}
+     * (report absence rather than a guess) and to refuse a {@code resumeSessionId} spawn that
+     * cannot be resolved reliably going forward.
+     */
+    static boolean isProvisionedWorktree(String cwd) {
+        return cwd != null && !cwd.isBlank() && Files.isRegularFile(Path.of(cwd, ".git"));
+    }
+
     // --- profile surface -----------------------------------------------------------------------
 
     /** The configured peer profile names (what {@code spawn(profile)} accepts). */

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

@@ -672,13 +672,29 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
     /** Add lazy on-disk session discovery to the base handle. */
     @Override
     public PeerHandle spawn(SpawnRequest req) {
+        String cwd = effectiveCwd(req);
+        // fleetd #249: refuse rather than silently resume into unverifiable territory. opencode's
+        // `-s <id>` flag itself resumes precisely — the resolved id is what fails, not the resume —
+        // but resolvedSessionId() below can never confirm (or later re-report) this handle's own
+        // identity without a fleetd-provisioned worktree (isProvisionedWorktree(cwd)), because the
+        // directory is shared and sessionIdForDirectory's "most recently updated row" heuristic can
+        // pick a sibling's session. Refusing here, before anything spawns, beats letting the member
+        // start and only then discovering fleetd can never again verify who it actually is.
+        if (req.resumeSessionId() != null && !req.resumeSessionId().isBlank()
+                && !isProvisionedWorktree(cwd)) {
+            throw new IllegalArgumentException("resumeSessionId requires a fleetd-provisioned "
+                    + "worktree for an opencode profile — without one, this member's cwd is shared "
+                    + "with other sessions, so fleetd can never reliably confirm (now or later) which "
+                    + "conversation it is actually running (fleetd #249). Pass fleet_spawn{worktree:"
+                    + "<ticket-slug>} to resume this member.");
+        }
         PeerHandle inner = super.spawn(req);
         // fleetd #175: the same profile config buildLaunch resolved for this spawn (requireProfile
         // is deterministic on req.profileName(), so re-resolving here costs a map lookup, not a
         // second decision) — SessionAwareHandle needs cfg.model() to know what THIS session should
         // be running.
         FleetConfig.Profile cfg = requireProfile(req.profileName());
-        return new SessionAwareHandle(inner, discovery, effectiveCwd(req), cfg,
+        return new SessionAwareHandle(inner, discovery, cwd, cfg,
                 this::memberHerdrSocketConfigured, discoveryUnavailableWarned, exhaustionSink);
     }
 
@@ -718,6 +734,17 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
          * the directory right now."
          */
         private final AtomicReference<String> resolvedSessionId = new AtomicReference<>();
+        /**
+         * fleetd #249: whether {@link #cwd} is a fleetd-provisioned git worktree
+         * ({@link HerdrPeerLauncher#isProvisionedWorktree}), computed once at spawn time since
+         * {@code cwd} never changes for this handle. When {@code false} the directory is shared
+         * with other sessions (the default no-worktree spawn inherits the lead's own cwd), so
+         * {@link OpenCodeSessionDiscovery#sessionIdForDirectory}'s "most recently updated row for
+         * this directory" heuristic can and does pick another session's row — see that class's
+         * javadoc. {@link #agentSessionId()} refuses to guess in that case: it reports absent
+         * rather than a possibly-foreign id.
+         */
+        private final boolean worktreeProvisioned;
 
         SessionAwareHandle(PeerHandle delegate, OpenCodeSessionDiscovery discovery, String cwd,
                             FleetConfig.Profile cfg,
@@ -731,6 +758,7 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
             this.discoveryUnavailable = discoveryUnavailable;
             this.discoveryUnavailableWarned = discoveryUnavailableWarned;
             this.exhaustionSink = exhaustionSink;
+            this.worktreeProvisioned = isProvisionedWorktree(cwd);
         }
 
         @Override
@@ -760,6 +788,9 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
             // built from (see OpenCodeLauncher#defaultDiscoveryRoot's javadoc for the full
             // reasoning). Scanning fleetd's own $HOME under that config would only ever find "no
             // row" and read as "resume unsupported" — declare it unavailable instead, once, loudly.
+            // Checked before the fleetd #249 worktree gate below: this OS-user mismatch makes
+            // discovery unusable regardless of whether cwd happens to be a provisioned worktree, so
+            // it earns the one-time WARN either way.
             if (discoveryUnavailable.getAsBoolean()) {
                 if (discoveryUnavailableWarned.compareAndSet(false, true)) {
                     log.warn("opencode session discovery unavailable: memberHerdrSocket is "
@@ -771,6 +802,16 @@ public final class OpenCodeLauncher extends HerdrPeerLauncher {
                 }
                 return null;
             }
+            // fleetd #249: cwd is shared with other sessions unless fleetd itself provisioned this
+            // worktree, and sessionIdForDirectory's directory-keyed heuristic cannot tell this
+            // 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.
+            if (!worktreeProvisioned) {
+                return null;
+            }
             // fleetd #234: once resolved, stay resolved. Re-deriving from `directory` on every call
             // would let this handle's identity drift to a sibling session that later shares the
             // same cwd and writes a newer row — see resolvedSessionId's javadoc.

fleetd/src/main/java/dev/ltms/fleet/msg/MessageService.java

@@ -746,7 +746,7 @@ public final class MessageService {
                 if (task != null) {
                     asyncTasksByWaiter.put(reply, task);
                 }
-                TurnToken token = new TurnToken(target, reply);
+                TurnToken token = new TurnToken(target, reply, content);
                 // The send has won the lock; the accepted-delivery hook records delegator ownership
                 // here (CB-548). It runs BEFORE enqueue so a throwing hook — onAccepted is now a
                 // public callback — fails the send without queuing a message that would orphan.

fleetd/src/main/java/dev/ltms/fleet/msg/TurnToken.java

@@ -9,12 +9,19 @@ import java.util.concurrent.CompletableFuture;
 public final class TurnToken {
     private final String target;
     private final CompletableFuture<Rendezvous.Resolution> waiter;
+    private final String injectedText;
 
     public TurnToken(String target, CompletableFuture<Rendezvous.Resolution> waiter) {
+        this(target, waiter, null);
+    }
+
+    public TurnToken(String target, CompletableFuture<Rendezvous.Resolution> waiter, String injectedText) {
         this.target = target;
         this.waiter = waiter;
+        this.injectedText = injectedText;
     }
 
     public String target() { return target; }
     public CompletableFuture<Rendezvous.Resolution> waiter() { return waiter; }
+    public String injectedText() { return injectedText; }
 }

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

@@ -426,19 +426,75 @@ public final class GitWorktrees implements Worktrees {
      * neutralized copy from ever showing up as a local modification the worker might commit. A config
      * the repo does not carry is skipped silently — no stub is invented for a file the repo does not
      * have, and one missing file must never fail provisioning.
+     *
+     * <p>fleetd #134. The neutralization above is correct and stays unconditional — the defect was
+     * that it was invisible on both sides. Neither the daemon's own log nor the worker sitting in the
+     * worktree could tell a stub from the repo's real file: a real worker read a 3-byte {@code {}}
+     * where the repo's {@code opencode.json} is 30+ lines, and reported — truthfully from what it
+     * could see, and wrongly — that a mount key did not exist. Two fixes, aimed at two different
+     * readers:
+     * <ul>
+     *   <li>the daemon operator reads {@link #log}, so the summary below names the denominator, what
+     *       was neutralized, and why anything was not — the same shape {@code overlayParity} reports
+     *       its own copy in;
+     *   <li>the worker reads its own worktree, not the daemon's log, so the same fact is recorded a
+     *       second time in worktree-scoped git config ({@code fleet.neutralizedConfig} /
+     *       {@code fleet.neutralizedConfigNote}, readable with {@code git config --worktree --get-all
+     *       fleet.neutralizedConfig}) rather than as a file in the working tree. A working-tree file
+     *       would show up in {@code git status} for the worker to trip on or commit; worktree-scoped
+     *       config lives in {@code .git/worktrees/<nonce>/config.worktree} and can never appear there.
+     *       This reuses the exact mechanism {@link #configureEnvironmentCredentialHelper} and
+     *       {@link #configureHttpsUrlRewriteForSshOrigin} already use for other worktree-local state.
+     * </ul>
      */
     private void isolateToolSurface(String worktreePath) {
         Path root = Path.of(worktreePath).toAbsolutePath().normalize();
+        List<String> neutralized = new ArrayList<>();
+        List<String> skipped = new ArrayList<>();
         for (WorktreeHostileConfig cfg : WORKTREE_HOSTILE_CONFIGS) {
-            neutralize(root, worktreePath, cfg);
+            if (neutralize(root, worktreePath, cfg)) {
+                neutralized.add(cfg.file());
+            } else {
+                skipped.add(cfg.file() + " absent");
+            }
         }
+        String detail = neutralized.isEmpty() ? String.join(", ", skipped)
+                : skipped.isEmpty() ? String.join(", ", neutralized)
+                : String.join(", ", neutralized) + " (" + String.join(", ", skipped) + ")";
+        log.info("tool-surface isolation: neutralized {} of {} configs: {} — the worktree copy is a "
+                        + "stub, not the repo's file; edit the real file in the primary checkout instead",
+                neutralized.size(), WORKTREE_HOSTILE_CONFIGS.size(), detail);
+        recordNeutralizedConfigForWorker(worktreePath, neutralized);
     }
 
-    private void neutralize(Path root, String worktreePath, WorktreeHostileConfig cfg) {
+    /**
+     * The worker-readable half of fleetd #134: record which files were neutralized where the worker
+     * itself can read it, without a working-tree file that would show up in {@code git status}.
+     * Worktree-scoped git config is per-worktree, lives under {@code .git/worktrees/<nonce>/} rather
+     * than the working tree, and this repo already relies on the same mechanism (and the same
+     * {@code extensions.worktreeConfig} enablement) for the credential helper and the SSH→HTTPS
+     * rewrite — see {@link #configureEnvironmentCredentialHelper}.
+     */
+    private void recordNeutralizedConfigForWorker(String worktreePath, List<String> neutralized) {
+        if (neutralized.isEmpty()) {
+            return;
+        }
+        exec("git", "-C", worktreePath, "config", "extensions.worktreeConfig", "true");
+        for (String file : neutralized) {
+            exec("git", "-C", worktreePath, "config", "--worktree", "--add", "fleet.neutralizedConfig", file);
+        }
+        exec("git", "-C", worktreePath, "config", "--worktree", "fleet.neutralizedConfigNote",
+                "the worktree copy of each fleet.neutralizedConfig path is a stub, not the repo's "
+                        + "committed file; edit the real file from the primary checkout instead");
+    }
+
+    /** @return true if {@code cfg} was neutralized (present, or created because {@link
+     *  WorktreeHostileConfig#createIfAbsent()}); false if the repo does not carry it and it was
+     *  left alone. */
+    private boolean neutralize(Path root, String worktreePath, WorktreeHostileConfig cfg) {
         Path target = root.resolve(cfg.file());
         if (!Files.exists(target) && !cfg.createIfAbsent()) {
-            log.debug("{} absent in the worktree — skipping (repo does not carry it)", cfg.file());
-            return;
+            return false;
         }
         try {
             Files.writeString(target, cfg.stub());
@@ -449,7 +505,7 @@ public final class GitWorktrees implements Worktrees {
         if (isTracked(root, cfg.file())) {
             exec("git", "-C", worktreePath, "update-index", "--skip-worktree", cfg.file());
         }
-        log.debug("neutralized {} — worker tool surface is launcher-mounted only", cfg.file());
+        return true;
     }
 
     @Override
@@ -486,25 +542,46 @@ public final class GitWorktrees implements Worktrees {
         }
         Path srcRoot = Path.of(repoRoot).toAbsolutePath().normalize();
         Path dstRoot = Path.of(worktreePath).toAbsolutePath().normalize();
+        List<String> copied = new ArrayList<>();
+        List<String> skipped = new ArrayList<>();
+        List<String> neutralized = new ArrayList<>();
         for (String rel : overlay) {
             Path src = srcRoot.resolve(rel).normalize();
             if (!Files.exists(src)) {
-                log.debug("parity overlay source missing — skipping {}", rel);
+                skipped.add(rel + " absent");
                 continue;
             }
             Path dst = dstRoot.resolve(rel).normalize();
             try {
                 Files.createDirectories(dst.getParent());
                 Files.copy(src, dst, StandardCopyOption.REPLACE_EXISTING, StandardCopyOption.COPY_ATTRIBUTES);
-                log.debug("copied parity overlay {}", rel);
+                copied.add(rel);
             } catch (IOException e) {
                 throw new WorktreeException("cannot copy overlay " + rel + ": " + e.getMessage(), e);
             }
             if (isTracked(dstRoot, rel)) {
                 exec("git", "-C", worktreePath, "update-index", "--skip-worktree", rel);
-                log.debug("marked overlay --skip-worktree {}", rel);
+                neutralized.add(rel);
             }
         }
+        // CB-148 point 3: a bare count ("copied 1") hides which candidates were even considered — the
+        // same shape of under-reporting this repo has been bitten by before. Name the denominator
+        // (every configured candidate), what was actually copied, and — for anything not copied —
+        // why, so a spawn's overlay outcome is legible from the log alone, no filesystem dig required.
+        String detail = copied.isEmpty() ? String.join(", ", skipped)
+                : skipped.isEmpty() ? String.join(", ", copied)
+                : String.join(", ", copied) + " (" + String.join(", ", skipped) + ")";
+        log.info("parity overlay: copied {} of {} candidates: {}", copied.size(), overlay.size(), detail);
+        // CB-134: a neutralized tracked file is otherwise a silent trap — a worker edits it, git
+        // ignores the change with no error, and nothing anywhere said the file could not be
+        // committed from this worktree. Name every file marked --skip-worktree here, with the
+        // consequence stated in the message itself, rather than adding a worktree-local marker
+        // file: acceptance criterion 1 requires the worktree hold exactly the configured overlay
+        // set and nothing else, so an extra marker would itself violate the fix.
+        if (!neutralized.isEmpty()) {
+            log.info("parity overlay marked --skip-worktree (cannot be committed from this worktree): {}",
+                    String.join(", ", neutralized));
+        }
     }
 
     @Override

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

@@ -0,0 +1,59 @@
+package dev.ltms.fleet;
+
+import dev.ltms.fleet.inject.BackendErrorPatternLookup;
+import dev.ltms.fleet.peer.MemberRole;
+import dev.ltms.fleet.session.MemberSession;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import java.util.List;
+import java.util.Map;
+import java.util.regex.Pattern;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertNull;
+
+/**
+ * fleetd #248 / fleetd#201 Unit 5: {@link Fleetd#backendErrorPatternLookup} is the factory that
+ * replaced the local lambda {@code Fleetd.main} used to build {@code backendErrorPatterns} — one
+ * of the two arguments {@code CompletionResolver} lost cleanly (0 compile errors, every test still
+ * green) when this ticket's measurement dropped it alongside {@code backendErrorSink}. This class
+ * proves the factory's own behaviour; {@code FleetdCompletionResolverWiringTest} proves {@code
+ * main} still passes its result into {@code CompletionResolver}.
+ */
+class FleetdBackendErrorPatternLookupTest {
+
+    private static MemberSession session(String terminal, String profile) {
+        return new MemberSession("pane-" + terminal, terminal, profile, MemberRole.DEV,
+                "/cwd", null, 0L, 0L, 0, MemberSession.State.READY, null, null);
+    }
+
+    @Test
+    @DisplayName("a target on a profile with a configured pattern resolves to that pattern")
+    void configuredProfileResolves() {
+        Map<String, Pattern> byProfile = Map.of("terra", Pattern.compile("(?i)503"));
+        BackendErrorPatternLookup lookup =
+                Fleetd.backendErrorPatternLookup(() -> List.of(session("term1", "terra")), byProfile);
+
+        assertEquals("(?i)503", lookup.patternFor("term1").pattern());
+    }
+
+    @Test
+    @DisplayName("a target on a profile with no configured pattern resolves to null")
+    void unconfiguredProfileResolvesToNull() {
+        Map<String, Pattern> byProfile = Map.of("terra", Pattern.compile("x"));
+        BackendErrorPatternLookup lookup =
+                Fleetd.backendErrorPatternLookup(() -> List.of(session("term1", "sol")), byProfile);
+
+        assertNull(lookup.patternFor("term1"));
+    }
+
+    @Test
+    @DisplayName("an unknown target resolves to null")
+    void unknownTargetResolvesToNull() {
+        BackendErrorPatternLookup lookup =
+                Fleetd.backendErrorPatternLookup(List::of, Map.of("terra", Pattern.compile("x")));
+
+        assertNull(lookup.patternFor("term_stranger"));
+    }
+}

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

@@ -0,0 +1,229 @@
+package dev.ltms.fleet;
+
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import dev.ltms.fleet.config.FleetConfig;
+import dev.ltms.fleet.guard.SubscriptionGuard;
+import dev.ltms.fleet.herdr.AgentControl;
+import dev.ltms.fleet.herdr.FakeHerdr;
+import dev.ltms.fleet.herdr.HerdrClient;
+import dev.ltms.fleet.herdr.WorkspaceControl;
+import dev.ltms.fleet.inject.BackendErrorSink;
+import dev.ltms.fleet.member.ClaudeCodeLauncher;
+import dev.ltms.fleet.member.CompositePeerLauncher;
+import dev.ltms.fleet.mcp.PrimaryRegistry;
+import dev.ltms.fleet.msg.InMemoryReplyInbox;
+import dev.ltms.fleet.msg.ReplyPushLoop;
+import dev.ltms.fleet.peer.Capability;
+import dev.ltms.fleet.peer.PeerHandle;
+import dev.ltms.fleet.peer.PeerLauncher;
+import dev.ltms.fleet.peer.SpawnRequest;
+import dev.ltms.fleet.placement.BackendOutagePolicy;
+import dev.ltms.fleet.placement.BackendQuarantine;
+import dev.ltms.fleet.placement.PlacementPolicies;
+import dev.ltms.fleet.session.MemberSession;
+import dev.ltms.fleet.session.SessionManager;
+import org.junit.jupiter.api.AfterEach;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import java.util.ArrayList;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.concurrent.CopyOnWriteArrayList;
+import java.util.concurrent.CountDownLatch;
+import java.util.concurrent.Executors;
+import java.util.concurrent.ScheduledExecutorService;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicLong;
+import java.util.concurrent.atomic.AtomicReference;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #248 / fleetd#201 Unit 5: {@link Fleetd#backendErrorSink} is the factory that replaced
+ * the local lambda {@code Fleetd.main} used to build {@code backendErrorSink} — the other half of
+ * the pair this ticket's measurement dropped cleanly (0 compile errors, every test still green).
+ *
+ * <p>Before this ticket, the closest thing to coverage was {@code
+ * dev.ltms.fleet.inject.BackendOutageFlowTest}, whose own class doc said it "mirrors {@code
+ * Fleetd.main}'s {@code backendErrorSink} lambda line-for-line" — a hand-copy that proves itself,
+ * never that {@code main} still wires the real thing. This class exercises the actual production
+ * factory instead. {@code FleetdCompletionResolverWiringTest} proves {@code main} still passes its
+ * result into {@code CompletionResolver}.
+ */
+class FleetdBackendErrorSinkTest {
+
+    private final List<ScheduledExecutorService> schedulers = new ArrayList<>();
+
+    @AfterEach
+    void tearDown() {
+        schedulers.forEach(ScheduledExecutorService::shutdownNow);
+    }
+
+    private static FleetConfig.Profile stubWorker(String profile, String credentialId) {
+        return new FleetConfig.Profile(profile, "http://gx00.gw:8000", "coder",
+                null, "FLEETD_WORKER_TOKEN", List.of("claude"), "tab", "fleetd-workers",
+                "w #{n}", null, null, null, null, null, null, null, null, null,
+                null, null, credentialId, null);
+    }
+
+    private static Map<String, FleetConfig.Profile> orderedProfiles() {
+        Map<String, FleetConfig.Profile> m = new LinkedHashMap<>();
+        m.put("terra", stubWorker("terra", "shared-openai"));
+        m.put("sol", stubWorker("sol", "shared-openai"));
+        return m;
+    }
+
+    /** Minimal recording {@code HerdrClient} for the LEAD pane — mirrors ReplyPushLoopTest's own. */
+    private static final class RecordingLeadClient implements HerdrClient {
+        private static final ObjectMapper MAPPER = new ObjectMapper();
+        private final List<Object> prompts = new CopyOnWriteArrayList<>();
+        volatile CountDownLatch sendLatch = new CountDownLatch(1);
+
+        @Override
+        public JsonNode call(String method, Object params) {
+            if ("agent.get".equals(method)) {
+                return MAPPER.createObjectNode().set("agent", MAPPER.createObjectNode()
+                        .put("terminal_id", "term_primary").put("agent_status", "idle"));
+            }
+            if ("agent.prompt".equals(method)) {
+                prompts.add(params);
+                sendLatch.countDown();
+            }
+            return MAPPER.createObjectNode();
+        }
+
+        @Override
+        public void close() {
+        }
+
+        int sendCount() {
+            return prompts.size();
+        }
+    }
+
+    /** A {@link PeerLauncher} that never actually spawns — enough to construct a bare {@link SessionManager}. */
+    private static final class NeverSpawnsLauncher implements PeerLauncher {
+        @Override
+        public Set<Capability> capabilities() {
+            return Set.of();
+        }
+
+        @Override
+        public Set<Capability> capabilitiesFor(String profileName) {
+            return Set.of();
+        }
+
+        @Override
+        public PeerHandle spawn(SpawnRequest req) {
+            throw new UnsupportedOperationException("not reachable — this test never acquires a session");
+        }
+
+        @Override
+        public Set<String> profiles() {
+            return Set.of();
+        }
+
+        @Override
+        public String defaultProfile() {
+            return null;
+        }
+
+        @Override
+        public String effectiveCwd(SpawnRequest req) {
+            throw new UnsupportedOperationException("not reachable — this test never acquires a session");
+        }
+
+        @Override
+        public List<String> parityOverlay(String profileName) {
+            return List.of();
+        }
+
+        @Override
+        public List<?> list() {
+            return List.of();
+        }
+
+        @Override
+        public int reapOrphanWorkers() {
+            return 0;
+        }
+
+        @Override
+        public void stop(String id) {
+        }
+
+        @Override
+        public boolean clearContext(String id) {
+            return false;
+        }
+    }
+
+    @Test
+    @DisplayName("a target with no resolvable profile logs and returns without recording an incident (never throws)")
+    void unresolvableProfileDoesNotRecordOrThrow() {
+        SessionManager sessions = new SessionManager(new NeverSpawnsLauncher());
+        BackendOutagePolicy outagePolicy = new BackendOutagePolicy(() -> 0L);
+        RecordingLeadClient leadClient = new RecordingLeadClient();
+        InMemoryReplyInbox inbox = new InMemoryReplyInbox();
+        PrimaryRegistry registry = new PrimaryRegistry(null);
+        ScheduledExecutorService scheduler = Executors.newSingleThreadScheduledExecutor();
+        schedulers.add(scheduler);
+        ReplyPushLoop pushLoop = new ReplyPushLoop(registry, new AgentControl(leadClient), inbox, scheduler, 3, 50);
+
+        BackendErrorSink sink = Fleetd.backendErrorSink(sessions, Map::of, outagePolicy, () -> pushLoop);
+        sink.onBackendError("term_unmapped", "matched line", "503 Service Unavailable");
+
+        assertTrue(outagePolicy.remainingCoolOffSeconds("shared-openai").isEmpty(),
+                "no credential is ever resolvable here, so nothing must be recorded");
+    }
+
+    @Test
+    @DisplayName("two distinct targets classified through the real factory start an incident and cool the credential")
+    void twoDistinctTargetsStartAnIncident() throws Exception {
+        FakeHerdr herdr = new FakeHerdr()
+                .readText("⏺ 503 Service Unavailable: upstream credential rejected\n❯ ");
+        Map<String, FleetConfig.Profile> profiles = orderedProfiles();
+        AtomicLong clockNanos = new AtomicLong(0L);
+        BackendOutagePolicy outagePolicy = new BackendOutagePolicy(clockNanos::get);
+        ClaudeCodeLauncher adapter = new ClaudeCodeLauncher(new AgentControl(herdr),
+                new WorkspaceControl(herdr), new SubscriptionGuard(Set.of("gx00.gw")),
+                profiles, "terra", _ -> "tok");
+        CompositePeerLauncher workers = new CompositePeerLauncher(List.of(adapter), "terra", profiles,
+                PlacementPolicies.weighted(), _ -> 0, null, BackendQuarantine.none(), outagePolicy);
+        SessionManager sessions = new SessionManager(workers);
+        MemberSession s1 = sessions.acquire("terra", null, null, null);
+        MemberSession s2 = sessions.acquire("terra", null, null, null);
+
+        PrimaryRegistry registry = new PrimaryRegistry(null);
+        registry.recordDelegation(s1.terminalId(), "term_primary");
+        registry.recordDelegation(s2.terminalId(), "term_primary");
+        InMemoryReplyInbox inbox = new InMemoryReplyInbox();
+        inbox.own(s1.terminalId());
+        inbox.own(s2.terminalId());
+        RecordingLeadClient leadClient = new RecordingLeadClient();
+        ScheduledExecutorService scheduler = Executors.newSingleThreadScheduledExecutor();
+        schedulers.add(scheduler);
+        ReplyPushLoop pushLoop = new ReplyPushLoop(registry, new AgentControl(leadClient), inbox, scheduler, 3, 50);
+        AtomicReference<ReplyPushLoop> pushLoopRef = new AtomicReference<>(pushLoop);
+
+        // The exact object under test: Fleetd's real production factory, not a hand copy.
+        BackendErrorSink sink = Fleetd.backendErrorSink(sessions, () -> profiles, outagePolicy, pushLoopRef::get);
+
+        sink.onBackendError(s1.terminalId(), "matched line", "503 Service Unavailable");
+        assertTrue(outagePolicy.remainingCoolOffSeconds("shared-openai").isEmpty(),
+                "one distinct target must not start a cool-off");
+
+        sink.onBackendError(s2.terminalId(), "matched line", "503 Service Unavailable");
+
+        assertTrue(leadClient.sendLatch.await(3, TimeUnit.SECONDS),
+                "the second distinct target must cross the threshold and nudge the lead");
+        var remaining = outagePolicy.remainingCoolOffSeconds("shared-openai");
+        assertTrue(remaining.isPresent(), "two distinct targets must start a cool-off");
+        assertEquals(1, leadClient.sendCount());
+    }
+}

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

@@ -0,0 +1,89 @@
+package dev.ltms.fleet;
+
+import java.nio.file.Files;
+import java.nio.file.Path;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #248: this is the test that was actually missing. {@code Fleetd.main} builds its {@code
+ * CompletionResolver} from an 8-argument constructor, and the ticket's own measurement proved two
+ * ways to silently unwire it — both compiled with 0 errors and left every existing test green:
+ *
+ * <ul>
+ *   <li>replacing the worktree/branch argument (the 8th) with {@code _ -> null} — drops
+ *       fleetd#241's fallback-report location entirely;</li>
+ *   <li>replacing {@code backendErrorPatterns, backendErrorSink} (5th/6th) with {@code
+ *       BackendErrorPatternLookup.legacy(), BackendErrorSink.none()} — drops fleetd#201 Unit 5's
+ *       backend-error classification and cool-off entirely.</li>
+ * </ul>
+ *
+ * <p>Neither mutation could be caught by any test that constructs its own {@code
+ * CompletionResolver} (every test before this one did exactly that) or by a test of {@link
+ * Fleetd#worktreeBranchLookup}, {@link Fleetd#backendErrorPatternLookup}, or {@link
+ * Fleetd#backendErrorSink} in isolation (see {@code FleetdWorktreeBranchLookupTest}, {@code
+ * FleetdBackendErrorPatternLookupTest}, {@code FleetdBackendErrorSinkTest}) — those prove the
+ * factories work, never that {@code main} still calls them. This class is a plain source-text
+ * assertion on {@code Fleetd.java} — crude, but honest about what it checks, and it turns red the
+ * instant the wiring is dropped, mirroring the same fallback shape {@link
+ * FleetdFleetAppConstructionTest} already uses for a different constructor argument.
+ *
+ * <p><b>This test checks source text, not runtime behaviour.</b> It never constructs a {@code
+ * CompletionResolver} and never runs {@code main}.
+ */
+class FleetdCompletionResolverWiringTest {
+
+    private static String fleetdSource() throws Exception {
+        return Files.readString(Path.of("src/main/java/dev/ltms/fleet/Fleetd.java"));
+    }
+
+    @Test
+    @DisplayName("[SOURCE TEXT] CompletionResolver's construction call still names backendErrorPatterns and backendErrorSink")
+    void backendErrorArgumentsAreStillNamedAtTheCallSite() throws Exception {
+        String source = fleetdSource();
+        assertTrue(source.contains(
+                        "exhaustionSink, backendErrorPatterns, backendErrorSink, System::nanoTime,"),
+                "CompletionResolver's construction call must still pass backendErrorPatterns and "
+                        + "backendErrorSink as its 5th/6th arguments. Replacing them with "
+                        + "BackendErrorPatternLookup.legacy()/BackendErrorSink.none() (fleetd #248's measured "
+                        + "mutation) compiles with 0 errors and leaves every behavioural test green — this "
+                        + "source check is what must go red instead.");
+    }
+
+    @Test
+    @DisplayName("[SOURCE TEXT] CompletionResolver's construction call still passes worktreeBranchLookup(sessions::roster)")
+    void worktreeBranchLookupIsStillPassedAtTheCallSite() throws Exception {
+        String source = fleetdSource();
+        assertTrue(source.contains("worktreeBranchLookup(sessions::roster)"),
+                "CompletionResolver's construction call must still pass worktreeBranchLookup(sessions::roster) "
+                        + "as its 8th (last) argument. Replacing it with the inert `_ -> null` (fleetd #248's "
+                        + "other measured mutation) compiles with 0 errors and leaves every behavioural test "
+                        + "green — this source check is what must go red instead.");
+        assertFalse(source.contains("System::nanoTime,\n                _ -> null"),
+                "the worktree/branch argument must never regress to the inert `_ -> null` literal");
+    }
+
+    @Test
+    @DisplayName("[SOURCE TEXT] backendErrorPatterns is assigned from the extracted backendErrorPatternLookup(...) factory")
+    void backendErrorPatternsComesFromTheFactory() throws Exception {
+        String source = fleetdSource();
+        assertTrue(source.contains(
+                        "BackendErrorPatternLookup backendErrorPatterns = backendErrorPatternLookup(sessions::roster,"),
+                "backendErrorPatterns must be assigned from Fleetd.backendErrorPatternLookup(...), not an "
+                        + "inline lambda that a source check on the CompletionResolver call alone cannot see "
+                        + "through");
+    }
+
+    @Test
+    @DisplayName("[SOURCE TEXT] backendErrorSink is assigned from the extracted backendErrorSink(...) factory")
+    void backendErrorSinkComesFromTheFactory() throws Exception {
+        String source = fleetdSource();
+        assertTrue(source.contains(
+                        "BackendErrorSink backendErrorSink = backendErrorSink(sessions, () -> config.get().profiles(),"),
+                "backendErrorSink must be assigned from Fleetd.backendErrorSink(...), not an inline lambda "
+                        + "that a source check on the CompletionResolver call alone cannot see through");
+    }
+}

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

@@ -0,0 +1,64 @@
+package dev.ltms.fleet;
+
+import dev.ltms.fleet.inject.CompletionResolver;
+import dev.ltms.fleet.peer.MemberRole;
+import dev.ltms.fleet.session.MemberSession;
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import java.util.ArrayList;
+import java.util.List;
+import java.util.function.Function;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertNull;
+
+/**
+ * fleetd #248: {@link Fleetd#worktreeBranchLookup} is the factory that replaced the anonymous
+ * lambda {@code Fleetd.main} used to build inline, as the 8th (last) argument to {@code
+ * CompletionResolver}'s constructor. Before this ticket that argument was untestable wiring:
+ * replacing it with {@code _ -> null} compiled clean and every existing test stayed green, because
+ * every existing test builds its own {@code CompletionResolver} directly rather than going through
+ * {@code main}. This class proves the factory's own behaviour; {@code
+ * FleetdCompletionResolverWiringTest} proves {@code main} still passes it in.
+ */
+class FleetdWorktreeBranchLookupTest {
+
+    private static MemberSession session(String terminal, String worktree, String branch) {
+        return new MemberSession("pane-" + terminal, terminal, "terra", MemberRole.DEV,
+                "/cwd", null, 0L, 0L, 0, MemberSession.State.READY, worktree, branch);
+    }
+
+    @Test
+    @DisplayName("a known target resolves to its session's worktree and branch")
+    void knownTargetResolves() {
+        Function<String, CompletionResolver.WorktreeBranch> lookup =
+                Fleetd.worktreeBranchLookup(() -> List.of(session("term1", "/wt/worker_x", "worker/x")));
+
+        CompletionResolver.WorktreeBranch resolved = lookup.apply("term1");
+
+        assertEquals("/wt/worker_x", resolved.worktree());
+        assertEquals("worker/x", resolved.branch());
+    }
+
+    @Test
+    @DisplayName("an unknown target resolves to null, not a thrown exception")
+    void unknownTargetResolvesToNull() {
+        Function<String, CompletionResolver.WorktreeBranch> lookup =
+                Fleetd.worktreeBranchLookup(() -> List.of(session("term1", "/wt/worker_x", "worker/x")));
+
+        assertNull(lookup.apply("term_stranger"));
+    }
+
+    @Test
+    @DisplayName("the roster is read through the supplier on every call, not snapshotted")
+    void rosterIsReadThroughOnEveryCall() {
+        List<MemberSession> roster = new ArrayList<>();
+        Function<String, CompletionResolver.WorktreeBranch> lookup = Fleetd.worktreeBranchLookup(() -> roster);
+
+        assertNull(lookup.apply("term_late"));
+        roster.add(session("term_late", "/wt/late", "worker/late"));
+
+        assertEquals("/wt/late", lookup.apply("term_late").worktree());
+    }
+}

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

@@ -1863,7 +1863,7 @@ class FleetConfigTest {
     }
 
     @Test
-    void parityOverlayDefaultsToEnvFilesOnly(@TempDir Path dir) throws Exception {
+    void parityOverlayDefaultsToDotEnvOnly(@TempDir Path dir) throws Exception {
         Path f = dir.resolve("no-overlay.yaml");
         Files.writeString(f, """
                 bind:
@@ -1874,9 +1874,11 @@ class FleetConfigTest {
                 """);
 
         FleetConfig cfg = FleetConfig.load(f);
-        assertEquals(List.of(".env", ".envrc"),
+        assertEquals(List.of(".env"),
                 cfg.profiles().get("gx10").parityOverlay(),
-                "the default parity overlay is the env files; settings.local.json is no longer copied by default");
+                "the default parity overlay is .env only (CB-148): .envrc is executable shell that "
+                        + "direnv runs on every cd, so it is no longer copied by default; settings.local.json "
+                        + "is also not copied by default");
     }
 
     @Test
@@ -1897,6 +1899,25 @@ class FleetConfigTest {
                 "an operator's explicit list survives verbatim — the default only changes when unset");
     }
 
+    @Test
+    void parityOverlayExplicitEnvrcOptInStillWorks(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("explicit-envrc-overlay.yaml");
+        Files.writeString(f, """
+                bind:
+                  port: 8080
+                profiles:
+                  gx10:
+                    baseUrl: http://gx10.gw:8000
+                    parityOverlay: [".env", ".envrc"]
+                """);
+
+        FleetConfig cfg = FleetConfig.load(f);
+        assertEquals(List.of(".env", ".envrc"),
+                cfg.profiles().get("gx10").parityOverlay(),
+                "an operator can still opt into copying .envrc explicitly (CB-148); it is only "
+                        + "dropped from the unset default, not removed as a capability");
+    }
+
     // ── CB-542: subscription:true must not smuggle an unguarded endpoint via env: ───────────────
 
     @Test

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

@@ -49,6 +49,7 @@ public final class FakeHerdr implements HerdrClient {
     private int pinnedStarts = 0;   // how many upcoming agent.start calls report a fixed pane
     private String pinnedStartTerminal;
     private String pinnedStartPane;
+    private Runnable onAgentStart; // fires the instant agent.start is called — see onAgentStart(Runnable)
     private volatile int agentGetOkCalls = Integer.MAX_VALUE; // how many agent.get calls succeed first
     private volatile String agentGetFailCode = null; // error code every agent.get call after that reports
 
@@ -152,6 +153,18 @@ public final class FakeHerdr implements HerdrClient {
         return this;
     }
 
+    /**
+     * Run {@code hook} synchronously the instant an {@code agent.start} call reaches this fake —
+     * i.e. the instant the peer PROCESS would start against a real herdr daemon. A test uses this
+     * to assert something is already true at that exact point (rather than merely true once
+     * {@code spawn()} returns), e.g. fleetd #149's trust-dialog seed having already been written to
+     * disk before the process herdr would launch ever starts.
+     */
+    public FakeHerdr onAgentStart(Runnable hook) {
+        this.onAgentStart = hook;
+        return this;
+    }
+
 
     /**
      * Seed a named agent into {@code agent.list} (e.g. an orphaned worker for CB-117 reaper tests).
@@ -250,6 +263,9 @@ public final class FakeHerdr implements HerdrClient {
                 case "agent.read" -> mapper.readTree(mapper.writeValueAsString(
                         java.util.Map.of("type", "agent_read", "read", java.util.Map.of("text", readText))));
                 case "agent.start" -> {
+                    if (onAgentStart != null) {
+                        onAgentStart.run();
+                    }
                     // Protocol 19: kind and pane_id are required — reject like the real daemon.
                     java.util.Map<?, ?> p = params instanceof java.util.Map<?, ?> m ? m : java.util.Map.of();
                     for (String required : new String[]{"kind", "pane_id"}) {

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

@@ -196,6 +196,26 @@ class CompletionResolverTest {
         assertEquals("complete report", waiter.getNow(null).text());
     }
 
+    @Test
+    void suppressesABareEchoWithOnlyTuiChrome() {
+        String injected = "Load the implementer skill. You own fleetd #999. ".repeat(12);
+        String scrape = injected + "\nDev auto - GPT-5.6 Terra OpenAI";
+
+        assertTrue(CompletionResolver.echoesInjectedBrief(scrape, injected));
+    }
+
+    @Test
+    void pinsTheMaximumTuiChromeExcess() {
+        String injected = "a".repeat(CompletionResolver.ECHO_MIN_CHARS);
+        String underMargin = injected + "b".repeat(CompletionResolver.MAX_ECHO_EXCESS_CHARS);
+        String overMargin = injected + "b".repeat(CompletionResolver.MAX_ECHO_EXCESS_CHARS + 1);
+
+        assertTrue(CompletionResolver.echoesInjectedBrief(underMargin, injected),
+                "the configured excess itself remains an echoed brief");
+        assertFalse(CompletionResolver.echoesInjectedBrief(overMargin, injected),
+                "one character beyond the excess must preserve the scrape as a real report");
+    }
+
     @Test
     void resolvesSynchronouslyBeforePostTurnContextClearing() {
         FakeHerdr herdr = new FakeHerdr().readText("⏺ previous answer\n❯ ");
@@ -748,19 +768,37 @@ class CompletionResolverTest {
     @Test
     void coverageIsOffWhenNoProfileHasAPatternConfigured() {
         assertEquals("off (no profile has an exhaustedPattern configured; profiles: [terra])",
-                CompletionResolver.coverage(Set.of("terra"), Set.of()));
+                CompletionResolver.coverage("exhaustedPattern", Set.of("terra"), Set.of()));
     }
 
     @Test
     void coverageIsFullWhenEveryProfileHasAPatternConfigured() {
         assertEquals("full (all profiles configured: [gx10, terra])",
-                CompletionResolver.coverage(Set.of("terra", "gx10"), Set.of("terra", "gx10")));
+                CompletionResolver.coverage("exhaustedPattern", Set.of("terra", "gx10"), Set.of("terra", "gx10")));
     }
 
     @Test
     void coverageIsPartialAndNamesWhichProfilesAreConfigured() {
         assertEquals("partial (configured: [terra]; not configured: [gx10])",
-                CompletionResolver.coverage(Set.of("terra", "gx10"), Set.of("terra")));
+                CompletionResolver.coverage("exhaustedPattern", Set.of("terra", "gx10"), Set.of("terra")));
+    }
+
+    /**
+     * Found live on 2026-09-03, reading a real boot log rather than a test. {@code coverage} is
+     * shared by two call sites — CB-578's {@code exhaustedPattern} line and fleetd#201 Unit 5's
+     * {@code errorPattern} line — but its "off" branch hard-coded the word {@code exhaustedPattern}.
+     * So a daemon with no {@code errorPattern} anywhere printed "no profile has an exhaustedPattern
+     * configured" directly beneath a line reporting that two profiles DO have one. Both lines were
+     * individually defensible and together they were nonsense, and the message sent an operator to
+     * set the wrong key.
+     *
+     * <p>Every earlier test here passed the exhaustion case only, so none of them could see it. This
+     * one pins that the message names the key the caller actually meant.
+     */
+    @Test
+    void coverageNamesTheConfigKeyItsCallerMeansRatherThanAlwaysSayingExhaustedPattern() {
+        assertEquals("off (no profile has an errorPattern configured; profiles: [gx10, terra])",
+                CompletionResolver.coverage("errorPattern", Set.of("terra", "gx10"), Set.of()));
     }
 
     // --- fleetd#201 Unit 1: target-keyed backend-error pattern + typed sink ----------------------

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

@@ -4,6 +4,9 @@ import ch.qos.logback.classic.Level;
 import ch.qos.logback.classic.Logger;
 import ch.qos.logback.classic.spi.ILoggingEvent;
 import ch.qos.logback.core.read.ListAppender;
+import com.fasterxml.jackson.databind.JsonNode;
+import com.fasterxml.jackson.databind.ObjectMapper;
+import com.fasterxml.jackson.databind.node.ObjectNode;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.guard.GuardException;
 import dev.ltms.fleet.guard.SubscriptionGuard;
@@ -23,11 +26,16 @@ import org.slf4j.LoggerFactory;
 import java.io.IOException;
 import java.nio.file.Files;
 import java.nio.file.Path;
+import java.nio.file.attribute.PosixFileAttributeView;
 import java.nio.file.attribute.PosixFilePermissions;
+import java.util.HashSet;
 import java.util.List;
 import java.util.Map;
 import java.util.Set;
 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.AtomicReference;
 import java.util.function.Function;
 import java.util.function.Supplier;
@@ -2098,4 +2106,433 @@ class ClaudeCodeLauncherTest {
         assertTrue(charterFile.getFileName().toString().startsWith("fleetd-role-charter-"),
                 "same file-naming scheme as before this fix (no wrapping directory): " + charterFile);
     }
+
+    // --- fleetd #149: workspace-trust dialog seed ------------------------------------------------
+    //
+    // Claude Code asks an interactive, un-timed "Is this a project you created or one you trust?"
+    // the first time it starts in a directory it has not seen. Every worktree: true spawn lands in
+    // a brand-new directory, so without a seed the member sits on that dialog forever — herdr still
+    // reports it healthy (agent_status: blocked, interactive_ready: true) — and never mounts the
+    // bridge MCP or calls fleet_reply. These tests start the REAL launcher (only the herdr transport
+    // is faked) so the seed is proven to run inside buildLaunch, before agent.start (the
+    // process-starting call) ever fires — a test that only checked the JSON writer in isolation
+    // would prove nothing about whether the launcher actually calls it at the right time.
+    //
+    // INCIDENT: the seed originally ran on ANY non-blank cwd. Running this file's own test suite —
+    // most of whose fixtures spawn with no cwd/configDir set, so both fall back to the real
+    // user.dir / ~/.claude.json — corrupted the operator's actual ~/.claude.json (it shrank from
+    // ~72 KB to a single seeded entry) the first time a manual mutation run made the write
+    // non-additive. The fix restricts the seed to isProvisionedWorktree(cwd) — a real .git FILE
+    // (not directory) — exactly the writeIdeOverlay gate already used for the same category of
+    // risk. Every test below marks its own worktree fixture with that .git file, and
+    // seedTrustDialogNeverWritesWhenCwdIsNotAProvisionedWorktree is the regression test for the
+    // incident itself.
+
+    private FleetConfig.Profile trustProfile(String configDir, String cwd) {
+        return new FleetConfig.Profile(
+                "ltms-local", "http://gx00.gw:8000", "coder", configDir, "FLEETD_WORKER_TOKEN",
+                List.of("claude"), "tab", "fleetd-workers", "w #{n}", "http://127.0.0.1:8765/mcp",
+                cwd, null);
+    }
+
+    /**
+     * Give {@code dir} the exact signature {@link ClaudeCodeLauncher#isProvisionedWorktree} (and
+     * {@code writeIdeOverlay} before it) checks for: a {@code .git} REGULAR FILE, never a
+     * directory. The content is never parsed by the trust seed, so any {@code gitdir:} pointer is
+     * fine.
+     */
+    private static void markAsProvisionedWorktree(Path dir) throws IOException {
+        Files.writeString(dir.resolve(".git"), "gitdir: /tmp/not-a-real-gitdir");
+    }
+
+    @Test
+    void seedsWorkspaceTrustForTheCwdBeforeTheProcessStarts(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        FakeHerdr herdr = new FakeHerdr();
+        Path claudeJson = configDir.resolve(".claude.json");
+        AtomicReference<Boolean> seededBeforeStart = new AtomicReference<>(false);
+        herdr.onAgentStart(() -> {
+            try {
+                if (!Files.exists(claudeJson)) {
+                    return;
+                }
+                JsonNode root = new ObjectMapper().readTree(claudeJson.toFile());
+                JsonNode project = root.path("projects").path(worktree.toString());
+                seededBeforeStart.set(project.path("hasTrustDialogAccepted").asBoolean(false)
+                        && project.path("hasCompletedProjectOnboarding").asBoolean(false));
+            } catch (IOException e) {
+                seededBeforeStart.set(false);
+            }
+        });
+
+        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();
+
+        assertTrue(herdr.called("agent.start"), "the hook must actually have fired during the spawn");
+        assertEquals(Boolean.TRUE, seededBeforeStart.get(),
+                "the trust entry for the cwd must already exist at the instant agent.start (the "
+                        + "process-starting herdr call) fires — not merely once spawn() returns");
+    }
+
+    @Test
+    void seedTrustDialogWritesBothTrustFlagsForTheResolvedCwd(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        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();
+
+        Path claudeJson = configDir.resolve(".claude.json");
+        assertTrue(Files.exists(claudeJson), "seeded into <configDir>/.claude.json");
+        JsonNode project = new ObjectMapper().readTree(claudeJson.toFile())
+                .path("projects").path(worktree.toString());
+        assertTrue(project.path("hasTrustDialogAccepted").asBoolean(false));
+        assertTrue(project.path("hasCompletedProjectOnboarding").asBoolean(false));
+    }
+
+    /**
+     * Criterion 3: seeding is additive. An existing {@code .claude.json} carries the operator's own
+     * project history and unrelated top-level settings — the seed must change only
+     * {@code projects.<cwd>} for THIS cwd and leave everything else, including a different project's
+     * own unrelated data, exactly as it was.
+     */
+    @Test
+    void seedTrustDialogIsAdditiveAndPreservesUnknownKeysAndOtherProjects(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        Path claudeJson = configDir.resolve(".claude.json");
+        Files.writeString(claudeJson, """
+                {
+                  "numStartups": 42,
+                  "oauthAccount": {"emailAddress": "operator@example.com"},
+                  "projects": {
+                    "/some/other/project": {
+                      "hasTrustDialogAccepted": true,
+                      "mcpServers": {"foo": {"type": "stdio", "command": "foo-mcp"}}
+                    }
+                  }
+                }
+                """);
+
+        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();
+
+        JsonNode root = new ObjectMapper().readTree(claudeJson.toFile());
+        assertEquals(42, root.path("numStartups").asInt(), "unrelated top-level key survives untouched");
+        assertEquals("operator@example.com", root.path("oauthAccount").path("emailAddress").asText(),
+                "an unrelated nested top-level key survives untouched");
+
+        JsonNode other = root.path("projects").path("/some/other/project");
+        assertTrue(other.path("hasTrustDialogAccepted").asBoolean(false),
+                "a different project's own trust entry survives");
+        assertEquals("foo-mcp", other.path("mcpServers").path("foo").path("command").asText(),
+                "a different project's own unrelated nested data survives");
+
+        JsonNode mine = root.path("projects").path(worktree.toString());
+        assertTrue(mine.path("hasTrustDialogAccepted").asBoolean(false));
+        assertTrue(mine.path("hasCompletedProjectOnboarding").asBoolean(false));
+    }
+
+    /**
+     * Criterion 2: where the profile sets no {@code configDir}, the seed goes to the default
+     * {@code ~/.claude.json}. {@code user.home} is redirected to a {@code @TempDir} for the
+     * duration of this test and restored in a {@code finally} — the real operator {@code
+     * ~/.claude.json} must never be touched by a test.
+     */
+    @Test
+    void seedTrustDialogTargetsDefaultClaudeJsonWhenConfigDirIsUnset(
+            @TempDir Path fakeHome, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        String originalHome = System.getProperty("user.home");
+        System.setProperty("user.home", fakeHome.toString());
+        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();
+
+            Path claudeJson = fakeHome.resolve(".claude.json");
+            assertTrue(Files.exists(claudeJson),
+                    "no configDir set — the default target is ~/.claude.json, here the redirected fake home");
+            JsonNode project = new ObjectMapper().readTree(claudeJson.toFile())
+                    .path("projects").path(worktree.toString());
+            assertTrue(project.path("hasTrustDialogAccepted").asBoolean(false));
+        } finally {
+            System.setProperty("user.home", originalHome);
+        }
+    }
+
+    /**
+     * Regression test for the fleetd #149 incident itself: a cwd that is NOT a fleetd-provisioned
+     * worktree (no {@code .git} FILE — the exact shape a real checkout, or an un-configured
+     * fallback cwd, has) must never be written to, however {@code configDir} is set. This is the
+     * fix for the exact defect that corrupted the operator's real {@code ~/.claude.json}.
+     */
+    @Test
+    void seedTrustDialogNeverWritesWhenCwdIsNotAProvisionedWorktree(
+            @TempDir Path configDir, @TempDir Path plainCwd) {
+        // plainCwd deliberately carries NO .git file — the same shape a real checkout's cwd
+        // fallback has.
+        FakeHerdr herdr = new FakeHerdr();
+        FleetConfig.Profile cfg = trustProfile(configDir.toString(), plainCwd.toString());
+        new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null)
+                .spawn();
+
+        assertFalse(Files.exists(configDir.resolve(".claude.json")),
+                "a non-worktree cwd must never get a .claude.json written for it — this is the fix "
+                        + "for the incident where writing unconditionally corrupted the operator's own "
+                        + "real ~/.claude.json via this file's own no-cwd/no-configDir test fixtures");
+    }
+
+    /**
+     * Same regression, for the default (no {@code configDir}) path — the exact combination (no
+     * {@code configDir}, no worktree-shaped {@code cwd}) that hit the operator's real
+     * {@code ~/.claude.json} during the incident. {@code user.home} is still redirected to a
+     * {@code @TempDir} out of caution, so even a reintroduced bug here cannot touch the real file.
+     */
+    @Test
+    void seedTrustDialogNeverWritesToDefaultHomeWhenCwdIsNotAProvisionedWorktree(
+            @TempDir Path fakeHome, @TempDir Path plainCwd) {
+        String originalHome = System.getProperty("user.home");
+        System.setProperty("user.home", fakeHome.toString());
+        try {
+            FakeHerdr herdr = new FakeHerdr();
+            FleetConfig.Profile cfg = trustProfile(null, plainCwd.toString());
+            new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
+                    new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null)
+                    .spawn();
+
+            assertFalse(Files.exists(fakeHome.resolve(".claude.json")),
+                    "the exact incident combination — no configDir, non-worktree cwd — must never "
+                            + "write, even to the (redirected) default ~/.claude.json");
+        } finally {
+            System.setProperty("user.home", originalHome);
+        }
+    }
+
+    // --- fleetd #149 review round 2: the write must be atomic and lock-protected -----------------
+    //
+    // The first round fixed WHICH cwd this can ever target. This round fixes HOW the target is
+    // written: .claude.json is large (tens of KB, dozens of projects on a real host), Claude Code
+    // itself rewrites it while running, and this daemon spawns several members in parallel as a
+    // matter of routine. The pre-fix Files.writeString(target, content) truncates target in place
+    // before writing the replacement bytes — a crash mid-write, or another writer's read landing in
+    // that window, loses data. Two more failure modes follow directly: (a) a crash/kill mid-write
+    // leaves target truncated, and (b) two concurrent spawns racing a naive read-modify-write let
+    // the second writer's write silently discard the first spawn's entry. The fix is
+    // ClaudeCodeLauncher.writeAtomically (sibling temp file + ATOMIC_MOVE, package-visible for the
+    // test below that proves it directly) plus TRUST_JSON_LOCK (a process-wide lock serialising
+    // every seedTrustDialog call this launcher itself makes).
+
+    /**
+     * Requested test 1: an existing, large-ish (not just a two-key fixture) .claude.json must never
+     * collapse. Asserts on the restored KEY SET — not merely that the result still parses as JSON,
+     * which the incident's 178-byte file also did.
+     */
+    @Test
+    void seedTrustDialogPreservesALargeExistingFileWithoutCollapsing(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        Path claudeJson = configDir.resolve(".claude.json");
+
+        ObjectMapper mapper = new ObjectMapper();
+        ObjectNode root = mapper.createObjectNode();
+        root.put("numStartups", 4200);
+        root.put("firstStartTime", "2025-01-01T00:00:00.000Z");
+        root.putObject("oauthAccount").put("emailAddress", "operator@example.com");
+        Set<String> otherProjectPaths = new HashSet<>();
+        ObjectNode projects = root.putObject("projects");
+        for (int i = 0; i < 30; i++) {
+            String path = "/Users/operator/code/project-" + i;
+            otherProjectPaths.add(path);
+            ObjectNode project = projects.putObject(path);
+            project.put("hasTrustDialogAccepted", true);
+            project.putObject("mcpServers").putObject("server-" + i).put("command", "server-" + i + "-mcp");
+        }
+        String before = mapper.writerWithDefaultPrettyPrinter().writeValueAsString(root);
+        Files.writeString(claudeJson, before);
+        long sizeBefore = Files.size(claudeJson);
+        assertTrue(sizeBefore > 4096,
+                "fixture must actually be large-ish to be a meaningful proof: " + sizeBefore + " bytes");
+
+        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();
+
+        long sizeAfter = Files.size(claudeJson);
+        assertTrue(sizeAfter >= sizeBefore,
+                "the file must never collapse below its pre-seed size — before=" + sizeBefore
+                        + " after=" + sizeAfter + " bytes (the incident shrank ~72 KB to 178 bytes)");
+
+        JsonNode after = mapper.readTree(claudeJson.toFile());
+        assertEquals(4200, after.path("numStartups").asInt(), "unrelated top-level key survives");
+        assertEquals("operator@example.com", after.path("oauthAccount").path("emailAddress").asText());
+        JsonNode afterProjects = after.path("projects");
+        for (String path : otherProjectPaths) {
+            assertTrue(afterProjects.path(path).path("hasTrustDialogAccepted").asBoolean(false),
+                    "pre-existing project entry " + path + " must survive");
+        }
+        assertEquals(otherProjectPaths.size() + 1, afterProjects.size(),
+                "exactly one NEW project entry (this worktree's) must be added, none dropped");
+        assertTrue(afterProjects.path(worktree.toString()).path("hasTrustDialogAccepted").asBoolean(false));
+    }
+
+    /**
+     * Requested test 2: two concurrent spawns for DIFFERENT cwd values sharing one configDir must
+     * both end up present in the final file — a naive concurrent read-modify-write would let the
+     * second writer's read (taken before the first writer's write lands) silently discard the
+     * first. A {@link CountDownLatch} — not a sleep — lines both threads up at the starting line so
+     * this does not depend on scheduling luck to be meaningful.
+     *
+     * <p>This is a test of {@code TRUST_JSON_LOCK}, not of {@code writeAtomically}: the lock fully
+     * serialises every {@code seedTrustDialog} call this launcher itself makes, so this test would
+     * pass even without atomicity. See {@link #writeAtomicallyNeverExposesATornFileToAConcurrentReader}
+     * for the test that exercises atomicity specifically.
+     */
+    @Test
+    void concurrentSeedsForDifferentCwdsBothSurvive(
+            @TempDir Path configDir, @TempDir Path worktreeA, @TempDir Path worktreeB) throws Exception {
+        markAsProvisionedWorktree(worktreeA);
+        markAsProvisionedWorktree(worktreeB);
+
+        CountDownLatch ready = new CountDownLatch(2);
+        CountDownLatch go = new CountDownLatch(1);
+        AtomicReference<Exception> failureA = new AtomicReference<>();
+        AtomicReference<Exception> failureB = new AtomicReference<>();
+
+        Thread ta = new Thread(spawnTask(configDir, worktreeA, ready, go, failureA), "spawn-a");
+        Thread tb = new Thread(spawnTask(configDir, worktreeB, ready, go, failureB), "spawn-b");
+        ta.start();
+        tb.start();
+
+        assertTrue(ready.await(5, TimeUnit.SECONDS), "both threads must reach the starting line");
+        go.countDown();
+        ta.join(5000);
+        tb.join(5000);
+        assertFalse(ta.isAlive(), "spawn A must finish within the timeout");
+        assertFalse(tb.isAlive(), "spawn B must finish within the timeout");
+        assertNull(failureA.get(), "spawn A must not throw: " + failureA.get());
+        assertNull(failureB.get(), "spawn B must not throw: " + failureB.get());
+
+        JsonNode root = new ObjectMapper().readTree(configDir.resolve(".claude.json").toFile());
+        assertTrue(root.path("projects").path(worktreeA.toString())
+                        .path("hasTrustDialogAccepted").asBoolean(false),
+                "worktree A's entry must survive the race");
+        assertTrue(root.path("projects").path(worktreeB.toString())
+                        .path("hasTrustDialogAccepted").asBoolean(false),
+                "worktree B's entry must survive the race");
+    }
+
+    private Runnable spawnTask(Path configDir, Path worktree, CountDownLatch ready, CountDownLatch go,
+                                AtomicReference<Exception> failure) {
+        return () -> {
+            try {
+                FakeHerdr herdr = new FakeHerdr();
+                FleetConfig.Profile cfg = trustProfile(configDir.toString(), worktree.toString());
+                ClaudeCodeLauncher launcher = new ClaudeCodeLauncher(new AgentControl(herdr),
+                        new WorkspaceControl(herdr), new SubscriptionGuard(Set.of("gx00.gw")),
+                        Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null);
+                ready.countDown();
+                go.await();
+                launcher.spawn();
+            } catch (Exception e) {
+                failure.set(e);
+            }
+        };
+    }
+
+    /**
+     * Requested test 3: the atomic write must preserve {@code .claude.json}'s existing {@code 0600}
+     * permissions, not silently widen them via a fresh temp file's own defaults landing on top of a
+     * file that had different (e.g. group-readable) permissions. Skips rather than fails on a
+     * filesystem with no POSIX permissions (e.g. Windows) — the same {@code assumeTrue} pattern this
+     * file already uses for {@code memberHerdrSocketWithWorktreeRootAndGroupPutsCharterUnderWorktreeRootAndSharesIt}.
+     */
+    @Test
+    void seedTrustDialogPreservesExisting0600Permissions(
+            @TempDir Path configDir, @TempDir Path worktree) throws Exception {
+        markAsProvisionedWorktree(worktree);
+        Path claudeJson = configDir.resolve(".claude.json");
+        Files.writeString(claudeJson, "{}");
+        assumeTrue(Files.getFileAttributeView(claudeJson, PosixFileAttributeView.class) != null,
+                "no POSIX permissions on this filesystem — skipping rather than failing");
+        Files.setPosixFilePermissions(claudeJson, PosixFilePermissions.fromString("rw-------"));
+
+        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();
+
+        assertEquals("rw-------", PosixFilePermissions.toString(Files.getPosixFilePermissions(claudeJson)),
+                "the atomic write must preserve .claude.json's existing 0600 permissions, not widen "
+                        + "them via a fresh temp file's own defaults");
+    }
+
+    /**
+     * Direct proof that {@link ClaudeCodeLauncher#writeAtomically} — not {@code TRUST_JSON_LOCK} —
+     * is what keeps a concurrent reader of {@code target} from ever observing a truncated or
+     * partially-written file. Calls {@code writeAtomically} directly (package-visible for exactly
+     * this test) rather than going through {@code seedTrustDialog}/{@code spawn()}, because {@link
+     * #concurrentSeedsForDifferentCwdsBothSurvive} above is protected by the lock and would pass
+     * even without atomicity — it proves the lock, not the atomic move. This test proves the atomic
+     * move specifically: a reader racing a writer OUTSIDE that lock (a second daemon process, or the
+     * operator's own live Claude Code — exactly what the lock cannot reach) must still never see a
+     * torn file.
+     *
+     * <p>The new content is made large (tens of MB) so a naive truncate-then-write has a real,
+     * non-instantaneous window for the busy-poll reader thread to land in — this is inherently a
+     * race, not a guaranteed-deterministic assertion, but it uses no {@code Thread.sleep} and
+     * reliably reproduced the torn read when run against the pre-fix
+     * {@code Files.writeString(target, content)} implementation (see the PR's mutation table).
+     */
+    @Test
+    void writeAtomicallyNeverExposesATornFileToAConcurrentReader(@TempDir Path dir) throws Exception {
+        Path target = dir.resolve(".claude.json");
+        String oldContent = "{\"marker\":\"OLD\"}";
+        Files.writeString(target, oldContent);
+        String newContent = "{\"marker\":\"NEW\",\"pad\":\"" + "x".repeat(20_000_000) + "\"}";
+
+        AtomicReference<String> tornRead = new AtomicReference<>();
+        AtomicBoolean stop = new AtomicBoolean(false);
+        Thread reader = new Thread(() -> {
+            while (!stop.get()) {
+                try {
+                    String seen = Files.readString(target);
+                    if (!seen.equals(oldContent) && !seen.equals(newContent)) {
+                        tornRead.compareAndSet(null, "torn read of length " + seen.length() + ": "
+                                + seen.substring(0, Math.min(seen.length(), 80)));
+                        stop.set(true);
+                    }
+                } catch (IOException ignored) {
+                    // ATOMIC_MOVE guarantees the path always resolves to old-or-new content once
+                    // readable at all — a transient "briefly missing during the rename" is fine to
+                    // ignore and keep sampling.
+                }
+            }
+        });
+        reader.start();
+        try {
+            ClaudeCodeLauncher.writeAtomically(target, newContent);
+        } finally {
+            stop.set(true);
+            reader.join(5000);
+        }
+
+        assertNull(tornRead.get(), "a concurrent reader must never observe a partially-written file: "
+                + tornRead.get());
+        assertEquals(newContent, Files.readString(target), "the final content must be the new content");
+    }
 }

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

@@ -323,11 +323,39 @@ class OpenCodeLauncherTest {
 
     // --- CB-547: resume + post-hoc session discovery --------------------------------------------
 
+    /**
+     * Give {@code dir} the exact signature {@link HerdrPeerLauncher#isProvisionedWorktree} checks
+     * for: a {@code .git} REGULAR FILE, never a directory. Content is never parsed by that gate, so
+     * any {@code gitdir:} pointer is fine. Mirrors {@code ClaudeCodeLauncherTest}'s helper of the
+     * same shape (fleetd #249).
+     */
+    private static void markAsProvisionedWorktree(Path dir) throws IOException {
+        Files.writeString(dir.resolve(".git"), "gitdir: /tmp/not-a-real-gitdir");
+    }
+
+    /**
+     * A fresh subdirectory of {@code configRoot}, marked as a provisioned worktree (fleetd #249),
+     * for tests that predate this gate and stood in a bare {@code "/work/dir"} string as their
+     * member's cwd — a directory that never existed on disk and, post-#249, would never pass
+     * {@link HerdrPeerLauncher#isProvisionedWorktree} either. Those tests are about the model
+     * mismatch / late-resolve machinery (fleetd #175/#234/#209), not about the worktree gate
+     * itself, so they need a cwd the gate accepts without changing what each test demonstrates.
+     */
+    private static String provisionedWorkDir(Path configRoot) throws IOException {
+        Path dir = Files.createDirectories(configRoot.resolve("work-dir"));
+        markAsProvisionedWorktree(dir);
+        return dir.toString();
+    }
+
     @Test
-    void aResumeSpawnPassesTheSessionIdAsDashS(@TempDir Path root) {
+    void aResumeSpawnIntoAProvisionedWorktreePassesTheSessionIdAsDashS(@TempDir Path root,
+                                                                        @TempDir Path worktree)
+            throws Exception {
+        markAsProvisionedWorktree(worktree);
         FakeHerdr herdr = new FakeHerdr();
         service(herdr, root, opencodeCfg("google/gemini-2.5-pro", null, null))
-                .spawn(new SpawnRequest(null, null, null, null, "ses_41b79fc90ffeI9E8uZv6VprUn2"));
+                .spawn(new SpawnRequest(null, worktree.toString(), null, null,
+                        "ses_41b79fc90ffeI9E8uZv6VprUn2"));
 
         List<String> args = startArgs(herdr);
         int s = args.indexOf("-s");
@@ -336,6 +364,27 @@ class OpenCodeLauncherTest {
                 "the resume target id follows -s");
     }
 
+    /**
+     * fleetd #249 acceptance criterion 3: without a fleetd-provisioned worktree, the member's cwd
+     * is shared with other sessions, so fleetd can never reliably confirm (now or later via {@link
+     * OpenCodeSessionDiscovery}) which conversation it is actually running. Refuse the spawn itself
+     * rather than silently launching opencode's {@code -s <id>} into unverifiable territory.
+     */
+    @Test
+    void aResumeSpawnWithoutAProvisionedWorktreeIsRefused(@TempDir Path root) {
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = service(herdr, root,
+                opencodeCfg("google/gemini-2.5-pro", null, null));
+
+        IllegalArgumentException e = assertThrows(IllegalArgumentException.class, () ->
+                launcher.spawn(new SpawnRequest(null, null, null, null,
+                        "ses_41b79fc90ffeI9E8uZv6VprUn2")));
+
+        assertTrue(e.getMessage().contains("worktree"), e.getMessage());
+        assertFalse(herdr.called("agent.start"),
+                "the refusal must happen before anything spawns — no pane, no process");
+    }
+
     @Test
     void aFreshSpawnCarriesNoSessionFlag(@TempDir Path root) {
         FakeHerdr herdr = new FakeHerdr();
@@ -348,25 +397,59 @@ class OpenCodeLauncherTest {
 
     @Test
     void theHandleDiscoversTheSessionIdForTheWorkersCwdOnlyAfterItAppears(@TempDir Path root,
-                                                                          @TempDir Path discRoot)
+                                                                          @TempDir Path discRoot,
+                                                                          @TempDir Path worktree)
             throws Exception {
+        markAsProvisionedWorktree(worktree);
         FakeHerdr herdr = new FakeHerdr();
         OpenCodeLauncher launcher = new OpenCodeLauncher(new AgentControl(herdr),
                 new WorkspaceControl(herdr), Map.of("gemini", opencodeCfg(null, null, null)),
                 "gemini", _ -> null, 0, System::currentTimeMillis, () -> { }, root, discRoot);
 
-        PeerHandle handle = launcher.spawn(new SpawnRequest(null, "/work/dir", null));
+        PeerHandle handle = launcher.spawn(new SpawnRequest(null, worktree.toString(), null));
 
         // opencode writes the record only when the session is first persisted — the instant the
         // pane is ready it does not exist, so agentSessionId() is null (never a spawn failure).
         assertNull(handle.agentSessionId(), "no record yet → null, not a spawn-time block");
         // Once the record appears (here: same cwd), lazy discovery resolves it — the handle's
         // session id matches its own worktree, not another's.
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_resolved", "/work/dir", 1000L);
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_resolved", worktree.toString(), 1000L);
         assertEquals("ses_resolved", handle.agentSessionId(),
                 "agentSessionId() re-scans and picks up a record that has since been written");
     }
 
+    /**
+     * fleetd #249 acceptance criterion 1, exercised through the real caller path (the handle
+     * {@code fleet_list} actually reads), not {@link OpenCodeSessionDiscovery} directly. Without a
+     * fleetd-provisioned worktree the member's cwd is shared — the default no-worktree spawn
+     * inherits the lead's own long-lived cwd — so even once a matching row appears (here:
+     * simulating another profile's session that happens to share the directory) the handle must
+     * report absence rather than guess. Measured real-world case (2026-09-03): the row it would
+     * otherwise pick was three days old and belonged to a different profile.
+     */
+    @Test
+    void theHandleNeverReportsAnIdForANonProvisionedCwdEvenAfterARowAppears(@TempDir Path root,
+                                                                             @TempDir Path discRoot)
+            throws Exception {
+        FakeHerdr herdr = new FakeHerdr();
+        OpenCodeLauncher launcher = new OpenCodeLauncher(new AgentControl(herdr),
+                new WorkspaceControl(herdr), Map.of("gemini", opencodeCfg(null, null, null)),
+                "gemini", _ -> null, 0, System::currentTimeMillis, () -> { }, root, discRoot);
+        // No markAsProvisionedWorktree — this cwd has no .git file, the shared-cwd shape a
+        // no-worktree spawn (or a real checkout) actually has.
+        String sharedCwd = root.resolve("shared-cwd").toString();
+
+        PeerHandle handle = launcher.spawn(new SpawnRequest(null, sharedCwd, null));
+
+        assertNull(handle.agentSessionId(), "no record yet → null, same as the provisioned case");
+        // A row for this exact directory now appears — e.g. a sibling member, or a stale session
+        // from days earlier, sharing the same unprovisioned cwd.
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_someone_elses", sharedCwd, 1000L);
+        assertNull(handle.agentSessionId(),
+                "a non-provisioned cwd must NEVER report an id, even once a row for it exists — "
+                        + "the row could belong to any other session sharing this directory");
+    }
+
     @Test
     void foreignWorkerMatchesOpencodePrefixButNotClaude() {
         String nonce = "abc123";
@@ -920,6 +1003,7 @@ class OpenCodeLauncherTest {
     @Test
     void theRealSessionManagerLateResolvePathCatchesAModelMismatch(@TempDir Path configRoot,
                                                                      @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         FakeHerdr herdr = new FakeHerdr();
         // xf's real shape (fleetd #175): weight:80, model "opencode/nemotron-3-ultra-free", no
         // credentialId — the profile that actually escaped the fleet's accounting.
@@ -929,7 +1013,7 @@ class OpenCodeLauncherTest {
         OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, sink);
 
         SessionManager sessions = new SessionManager(launcher);
-        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+        MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
 
         // Real late-resolve path, driven BEFORE opencode has written its session row — same shape
         // as production the instant a pane goes ready.
@@ -940,7 +1024,7 @@ class OpenCodeLauncherTest {
 
         // opencode writes its row late, running gpt-5.6-sol (a PAID credential) instead of the
         // withdrawn free model the profile actually asked for — the exact fleetd #175 scenario.
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
 
         // Drive the SAME real late-resolve path again: sessions.get() -> resolveAgentSessionId ->
@@ -959,12 +1043,13 @@ class OpenCodeLauncherTest {
     @Test
     void aProviderPrefixedModelMatchingBothIdAndProviderIsNotAMismatch(@TempDir Path configRoot,
                                                                         @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                .spawn(new SpawnRequest(null, workDir, null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
 
         assertEquals("ses_x", handle.agentSessionId());
@@ -975,12 +1060,13 @@ class OpenCodeLauncherTest {
     @Test
     void aGxProviderPrefixedModelMatchingBothIdAndProviderIsNotAMismatch(@TempDir Path configRoot,
                                                                           @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("gx/deepseek-v4-flash", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                .spawn(new SpawnRequest(null, workDir, null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
 
         assertEquals("ses_x", handle.agentSessionId());
@@ -997,12 +1083,13 @@ class OpenCodeLauncherTest {
     @Test
     void aMissingProviderIdInTheEvidenceIsUnknownNotAMismatchWhenTheIdMatches(
             @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                .spawn(new SpawnRequest(null, workDir, null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-terra\"}");
 
         assertEquals("ses_x", handle.agentSessionId());
@@ -1018,12 +1105,13 @@ class OpenCodeLauncherTest {
     @Test
     void aMissingProviderIdInTheEvidenceStillCatchesARealIdMismatch(
             @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                .spawn(new SpawnRequest(null, workDir, null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-sol\"}");
 
         assertEquals("ses_x", handle.agentSessionId());
@@ -1042,12 +1130,13 @@ class OpenCodeLauncherTest {
     @Test
     void aBareModelWithNoProviderPrefixMatchesOnIdAloneAndIsNotAMismatch(@TempDir Path configRoot,
                                                                           @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("deepseek-v4-flash", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                .spawn(new SpawnRequest(null, workDir, null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
 
         assertEquals("ses_x", handle.agentSessionId());
@@ -1064,6 +1153,7 @@ class OpenCodeLauncherTest {
     @Test
     void aRealIdMismatchLogsAnErrorNamingBothModelsAndQuarantinesThroughTheSink(
             @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(target + "|" + reason);
         FleetConfig.Profile cfg = opencodeCfg("opencode/nemotron-3-ultra-free", null, null);
@@ -1075,8 +1165,8 @@ class OpenCodeLauncherTest {
         PeerHandle handle;
         try {
             handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                    .spawn(new SpawnRequest(null, "/work/dir", null));
-            OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                    .spawn(new SpawnRequest(null, workDir, null));
+            OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                     "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
             assertEquals("ses_x", handle.agentSessionId());
         } finally {
@@ -1111,11 +1201,12 @@ class OpenCodeLauncherTest {
     @Test
     void unknownOrUnparseableModelEvidenceNeverQuarantines(@TempDir Path configRoot,
                                                              @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
+                .spawn(new SpawnRequest(null, workDir, null));
 
         // No row yet at all.
         assertNull(handle.agentSessionId());
@@ -1137,12 +1228,13 @@ class OpenCodeLauncherTest {
     @Test
     void aProfileWithNoConfiguredModelIsNeverCheckedForAMismatch(@TempDir Path configRoot,
                                                                   @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg(null, null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+                .spawn(new SpawnRequest(null, workDir, null));
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"anything-at-all\",\"providerID\":\"anyone\"}");
 
         assertEquals("ses_x", handle.agentSessionId());
@@ -1166,21 +1258,22 @@ class OpenCodeLauncherTest {
     @Test
     void modelCheckReadsTheResolvedSessionsOwnRowNotWhateverIsNewestInTheSharedDirectory(
             @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         List<String> exhausted = new ArrayList<>();
         ExhaustionSink sink = (target, reason, profile) -> exhausted.add(reason);
         FleetConfig.Profile cfg = opencodeCfg("openai/gpt-5.6-terra", null, null);
         PeerHandle handle = serviceWithSink(new FakeHerdr(), configRoot, discRoot, cfg, sink)
-                .spawn(new SpawnRequest(null, "/work/dir", null));
+                .spawn(new SpawnRequest(null, workDir, null));
 
         // Our own session's row, correctly matching the profile's requested model.
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_ours", "/work/dir", 1000L,
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_ours", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-terra\",\"providerID\":\"openai\"}");
         assertEquals("ses_ours", handle.agentSessionId(), "resolves to our own session");
         assertTrue(exhausted.isEmpty(), "matching model → no mismatch on first resolve: " + exhausted);
 
         // A sibling member, spawned later into the SAME shared directory (no worktree, fleetd
         // #234's default), writes a newer row running a totally different model.
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_sibling", "/work/dir", 9000L,
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_sibling", workDir, 9000L,
                 "{\"id\":\"deepseek-v4-flash\",\"providerID\":\"gx\"}");
 
         assertEquals("ses_ours", handle.agentSessionId(),
@@ -1212,6 +1305,7 @@ class OpenCodeLauncherTest {
     @Test
     void aSpawnTimeModelMismatchActuallyQuarantinesTheCredentialThroughTheRealAcquirePath(
             @TempDir Path configRoot, @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         FleetConfig.Profile cfg = opencodeCfgWithCredential(
                 "terra", "opencode/nemotron-3-ultra-free", "openai-shared");
         Map<String, FleetConfig.Profile> profiles = Map.of(cfg.profile(), cfg);
@@ -1232,14 +1326,14 @@ class OpenCodeLauncherTest {
 
         // The mismatching row exists BEFORE the spawn — reproducing fleetd #234's exact timing:
         // opencode's session table already carries evidence by the moment acquire() first asks.
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
 
         assertFalse(quarantine.isQuarantined("openai-shared"), "nothing quarantined before the spawn");
 
         // The real production entrypoint: acquire() builds the MemberSession by calling
         // handle.agentSessionId() BEFORE registry.put() runs.
-        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+        MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
 
         assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
         assertTrue(quarantine.isQuarantined("openai-shared"),
@@ -1258,6 +1352,7 @@ class OpenCodeLauncherTest {
     @Test
     void aRosterOnlySinkSilentlyDropsTheSpawnTimeQuarantine(@TempDir Path configRoot,
                                                               @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         FleetConfig.Profile cfg = opencodeCfgWithCredential(
                 "terra", "opencode/nemotron-3-ultra-free", "openai-shared");
         BackendQuarantine quarantine = new BackendQuarantine(() -> 0L, TimeUnit.SECONDS.toNanos(1800));
@@ -1270,10 +1365,10 @@ class OpenCodeLauncherTest {
         OpenCodeLauncher launcher = serviceWithSink(herdr, configRoot, discRoot, cfg, rosterOnlySink);
         SessionManager sessions = new SessionManager(launcher);
 
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
 
-        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+        MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
 
         assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
         assertFalse(quarantine.isQuarantined("openai-shared"),
@@ -1301,6 +1396,7 @@ class OpenCodeLauncherTest {
     @Test
     void theSpawnTimeQuarantineSurvivesTheFleetdStyleForwardingHop(@TempDir Path configRoot,
                                                                      @TempDir Path discRoot) throws Exception {
+        String workDir = provisionedWorkDir(configRoot);
         FleetConfig.Profile cfg = opencodeCfgWithCredential(
                 "terra", "opencode/nemotron-3-ultra-free", "openai-shared");
         Map<String, FleetConfig.Profile> profiles = Map.of(cfg.profile(), cfg);
@@ -1332,12 +1428,12 @@ class OpenCodeLauncherTest {
         };
         exhaustionSinkRef.set(realSink);
 
-        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", "/work/dir", 1000L,
+        OpenCodeSessionDiscoveryTest.writeRecord(discRoot, "ses_x", workDir, 1000L,
                 "{\"id\":\"gpt-5.6-sol\",\"providerID\":\"openai\"}");
 
         assertFalse(quarantine.isQuarantined("openai-shared"), "nothing quarantined before the spawn");
 
-        MemberSession acquired = sessions.acquire(cfg.profile(), "/work/dir", null, null);
+        MemberSession acquired = sessions.acquire(cfg.profile(), workDir, null, null);
 
         assertEquals("ses_x", acquired.agentSessionId(), "the id itself still resolves correctly");
         assertTrue(quarantine.isQuarantined("openai-shared"),

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

@@ -56,7 +56,11 @@ class MessageServiceTest {
 
     /** Run {@code send} on a background thread; the current thread drives the worker's turn. */
     private CompletableFuture<MessageService.Reply> sendAsync() {
-        return CompletableFuture.supplyAsync(() -> messages.send(T, "do the task", 5000));
+        return sendAsync("do the task");
+    }
+
+    private CompletableFuture<MessageService.Reply> sendAsync(String content) {
+        return CompletableFuture.supplyAsync(() -> messages.send(T, content, 5000));
     }
 
     private void awaitWaiting() throws InterruptedException {
@@ -86,6 +90,94 @@ class MessageServiceTest {
         assertTrue(reply.completed(), "a scraped completion still counts as completed");
     }
 
+    @Test
+    void completionFallbackReplacesAnEchoedInjectedBriefWithNoReportOutcome() throws Exception {
+        String brief = "Implement the requested change. ".repeat(20);
+        CompletableFuture<MessageService.Reply> send = sendAsync(brief);
+        awaitWaiting();
+
+        herdr.readText("$ prompt");
+        injector.onStatus(T, AgentStatus.IDLE);
+        injector.onStatus(T, AgentStatus.WORKING);
+        herdr.readText("⏺ " + brief + "\n❯ ");
+        injector.onStatus(T, AgentStatus.IDLE);
+
+        MessageService.Reply reply = send.get(5, TimeUnit.SECONDS);
+        assertEquals(MessageService.Outcome.COMPLETED_UNREPLIED, reply.outcome());
+        assertEquals(CompletionResolver.NO_REPORT_PREFIX + "]", reply.text(),
+                "the real injector -> completion fallback path must not return the lead's brief");
+    }
+
+    @Test
+    void completionFallbackKeepsARealReportThatRestatesTheWholeBrief() throws Exception {
+        String brief = "Load the implementer skill. You own fleetd #999. ".repeat(12);
+        String report = brief + "\n\n## Report\n"
+                + ("I implemented the fix in GitWorktrees.java, added five tests, ran mvn clean install "
+                + "and got 1168 tests with 0 failures. Commit 321d8dc pushed. ").repeat(8);
+        CompletableFuture<MessageService.Reply> send = sendAsync(brief);
+        awaitWaiting();
+
+        herdr.readText("$ prompt");
+        injector.onStatus(T, AgentStatus.IDLE);
+        injector.onStatus(T, AgentStatus.WORKING);
+        herdr.readText("⏺ " + report + "\n❯ ");
+        injector.onStatus(T, AgentStatus.IDLE);
+
+        MessageService.Reply reply = send.get(5, TimeUnit.SECONDS);
+        assertEquals(MessageService.Outcome.COMPLETED_UNREPLIED, reply.outcome());
+        assertEquals(report.strip(), reply.text(), "a report that restates the whole brief must survive unchanged");
+    }
+
+    @Test
+    void completionFallbackNamesTheKnownWorktreeAndBranchForAnEchoedBrief() throws Exception {
+        FakeHerdr localHerdr = new FakeHerdr();
+        Rendezvous localRendezvous = new Rendezvous();
+        java.util.concurrent.atomic.AtomicLong localClock = new java.util.concurrent.atomic.AtomicLong();
+        CompletionResolver localCompletion = new CompletionResolver(new AgentControl(localHerdr), localRendezvous,
+                ExhaustedPatternLookup.none(), ExhaustionSink.none(),
+                dev.ltms.fleet.inject.BackendErrorPatternLookup.legacy(),
+                dev.ltms.fleet.inject.BackendErrorSink.none(),
+                () -> localClock.addAndGet(CompletionResolver.MIN_TURN_NANOS + 1),
+                target -> new CompletionResolver.WorktreeBranch("/tmp/member-worktree", "worker/cb241"));
+        Injector localInjector = new Injector(new AgentControl(localHerdr), localCompletion);
+        MessageService localMessages = new MessageService(new AgentControl(localHerdr), localInjector,
+                localRendezvous, new InMemoryReplyInbox());
+        String brief = "Implement the requested change. ".repeat(20);
+        CompletableFuture<MessageService.Reply> send =
+                CompletableFuture.supplyAsync(() -> localMessages.send(T, brief, 5000));
+        long deadline = System.currentTimeMillis() + 2000;
+        while (!localRendezvous.isWaiting(T) && System.currentTimeMillis() < deadline) {
+            Thread.sleep(5);
+        }
+        assertTrue(localRendezvous.isWaiting(T));
+
+        localHerdr.readText("$ prompt");
+        localInjector.onStatus(T, AgentStatus.IDLE);
+        localInjector.onStatus(T, AgentStatus.WORKING);
+        localHerdr.readText("⏺ " + brief + "\n❯ ");
+        localInjector.onStatus(T, AgentStatus.IDLE);
+
+        assertEquals(CompletionResolver.NO_REPORT_PREFIX + " worktree=/tmp/member-worktree "
+                        + "branch=worker/cb241]", send.get(5, TimeUnit.SECONDS).text());
+    }
+
+    @Test
+    void completionFallbackKeepsTheClippedMarkerWhenAnEchoedBriefIsTooLong() throws Exception {
+        String brief = "a".repeat(4_001); // CompletionResolver's 4,000-character scrape cap
+        CompletableFuture<MessageService.Reply> send = sendAsync(brief);
+        awaitWaiting();
+
+        herdr.readText("$ prompt");
+        injector.onStatus(T, AgentStatus.IDLE);
+        injector.onStatus(T, AgentStatus.WORKING);
+        herdr.readText("⏺ " + brief + "\n❯ ");
+        injector.onStatus(T, AgentStatus.IDLE);
+
+        assertEquals(CompletionResolver.NO_REPORT_PREFIX + "]\n"
+                        + "[Pane tail clipped: member did not call fleet_reply.]",
+                send.get(5, TimeUnit.SECONDS).text());
+    }
+
     @Test
     void backendErrorScrapeThroughMessageServiceFailsInsteadOfBecomingReplyText() throws Exception {
         // fleetd#164 (part 2 addendum): a scrape that reads cleanly but is only the backend's own

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

@@ -666,6 +666,103 @@ class GitWorktreesTest {
         assertEquals("", status(Path.of(wt), ".autoenv"), ".autoenv still shows as modified");
     }
 
+    // ---- fleetd #134: isolateToolSurface must report what it neutralized (to the daemon operator's
+    // log) and record it where the worker itself can read it (worktree-scoped git config), without
+    // ever showing up in the worker's own `git status`. All drive the real provisioning path,
+    // GitWorktrees#add, per criterion 5 — the whole provisioned worktree is what's under test here. ----
+
+    private static Path initRepoWithAllThreeConfigs(Path dir) throws Exception {
+        Files.createDirectories(dir);
+        git(dir, "init", "-q", "-b", "main");
+        git(dir, "config", "user.email", "test@example.invalid");
+        git(dir, "config", "user.name", "Test");
+        Files.writeString(dir.resolve(".mcp.json"), WITH_SERVERS);
+        Files.writeString(dir.resolve("opencode.json"), OPENCODE_WITH_FILE_REF);
+        Files.writeString(dir.resolve(".autoenv"), AUTOENV_WITH_DIRECTIVE);
+        Files.writeString(dir.resolve("README.md"), "seed\n");
+        git(dir, "add", ".mcp.json", "opencode.json", ".autoenv", "README.md");
+        git(dir, "commit", "-q", "-m", "seed");
+        return dir;
+    }
+
+    /** Criterion 1, all three present: the summary names the denominator and every neutralized file. */
+    @Test
+    void isolateToolSurfaceLogsAllThreeConfigsNeutralized(@TempDir Path tmp) throws Exception {
+        reportingLogger.setLevel(Level.INFO);
+        Path repo = initRepoWithAllThreeConfigs(tmp.resolve("repo"));
+
+        new GitWorktrees(tmp.resolve("wts").toString()).add(repo.toString(), "cb-134-log-all", "HEAD");
+
+        assertTrue(capturedMessages().contains(
+                        "tool-surface isolation: neutralized 3 of 3 configs: .mcp.json, opencode.json, "
+                                + ".autoenv — the worktree copy is a stub, not the repo's file; edit the "
+                                + "real file in the primary checkout instead"),
+                "expected the all-neutralized summary line, got:\n" + capturedMessages());
+    }
+
+    /** Criterion 1, two absent: the summary must still name the denominator and say why. */
+    @Test
+    void isolateToolSurfaceLogsAbsentConfigsWithReason(@TempDir Path tmp) throws Exception {
+        reportingLogger.setLevel(Level.INFO);
+        Path repo = initRepo(tmp.resolve("repo")); // only .mcp.json + README committed
+
+        new GitWorktrees(tmp.resolve("wts").toString()).add(repo.toString(), "cb-134-log-partial", "HEAD");
+
+        assertTrue(capturedMessages().contains(
+                        "tool-surface isolation: neutralized 1 of 3 configs: .mcp.json (opencode.json "
+                                + "absent, .autoenv absent) — the worktree copy is a stub, not the repo's "
+                                + "file; edit the real file in the primary checkout instead"),
+                "expected the partial summary line, got:\n" + capturedMessages());
+    }
+
+    /**
+     * Criterion 2. The daemon's own log is invisible to the worker process — it never reads fleetd's
+     * stdout. This is the mechanism the worker itself can query, from inside its own worktree, to
+     * learn "this file is neutralized here, the repo's real file differs" instead of trusting what it
+     * just read on disk.
+     */
+    @Test
+    void aWorkerCanDiscoverNeutralizedConfigsFromWorktreeScopedGitConfig(@TempDir Path tmp) throws Exception {
+        Path repo = initRepoWithAllThreeConfigs(tmp.resolve("repo"));
+
+        String wt = new GitWorktrees(tmp.resolve("wts").toString())
+                .add(repo.toString(), "cb-134-discover", "HEAD");
+
+        String recorded = gitOutput(Path.of(wt), "config", "--worktree", "--get-all", "fleet.neutralizedConfig");
+        Set<String> files = new HashSet<>();
+        for (String line : recorded.split("\\R")) {
+            if (!line.isBlank()) {
+                files.add(line.trim());
+            }
+        }
+        assertEquals(Set.of(".mcp.json", "opencode.json", ".autoenv"), files,
+                "the worker-readable record must name every neutralized file: " + recorded);
+
+        String note = gitOutput(Path.of(wt), "config", "--worktree", "--get", "fleet.neutralizedConfigNote").trim();
+        assertTrue(note.contains("stub"), "note must say the worktree copy is a stub: " + note);
+        assertTrue(note.contains("primary checkout"),
+                "note must state the consequence — where to edit the real file instead: " + note);
+    }
+
+    /**
+     * Criterion 3. Whatever fleetd#134's worker-discovery mechanism writes must never appear as
+     * untracked or modified in the worker's own `git status` — a worker that sees a stray file either
+     * commits it by mistake or burns a turn asking about it. This runs the full porcelain status, not
+     * a single-file check, so any leftover file anywhere in the worktree would fail it.
+     */
+    @Test
+    void aProvisionedWorktreeHasCleanGitStatusDespiteNeutralizedConfigRecordkeeping(@TempDir Path tmp)
+            throws Exception {
+        Path repo = initRepoWithAllThreeConfigs(tmp.resolve("repo"));
+
+        String wt = new GitWorktrees(tmp.resolve("wts").toString())
+                .add(repo.toString(), "cb-134-clean-status", "HEAD");
+
+        assertEquals("", fullStatus(Path.of(wt)),
+                "a freshly provisioned worktree must show a clean `git status --porcelain`, including "
+                        + "after the worker-readable neutralized-config record was written");
+    }
+
     /**
      * CB-578 stage C, acceptance criterion 1. A dirty worktree — a tracked edit plus a brand-new
      * untracked file, exactly the shape lost in CB-576 — must land in {@code refs/wip/<branch>}'s
@@ -1177,4 +1274,112 @@ class GitWorktreesTest {
                     + "the refusal must happen before `git worktree add` ever runs");
         }
     }
+
+    // ---- fleetd #134 / #148 point 3: overlayParity must report what it did, and marking a tracked
+    // file --skip-worktree must say the file can no longer be committed from this worktree. Drives
+    // overlayParity directly against a real worktree (git worktree add, no GitWorktrees#add) so these
+    // tests are independent of origin/credential-helper provisioning, which is not under test here. ----
+
+    /** A bare worktree, sibling to {@code repo}, created with plain git — the target overlayParity
+     *  copies into. Deliberately not {@link GitWorktrees#add}: that method does unrelated
+     *  provisioning (origin rewrite, .mcp.json neutralization, credential helper) that would only
+     *  add noise to the log assertions below. */
+    private static Path bareWorktree(Path repo, Path wtDir, String branch) throws Exception {
+        git(repo, "worktree", "add", "-q", wtDir.toString(), "-b", branch, "HEAD");
+        return wtDir;
+    }
+
+    /** Acceptance criterion 1: only the configured candidates land in the worktree — nothing else
+     *  from the source tree leaks in alongside them. */
+    @Test
+    void overlayParityCopiesExactlyTheConfiguredFilesAndNothingElse(@TempDir Path tmp) throws Exception {
+        Path repo = initRepo(tmp.resolve("repo"));
+        Files.writeString(repo.resolve(".env"), "A=1\n");
+        Files.writeString(repo.resolve(".envrc"), "export A=1\n");
+        Files.writeString(repo.resolve("not-overlaid.txt"), "must not be copied\n");
+        Path wt = bareWorktree(repo, tmp.resolve("wt"), "cb134-exact");
+
+        new GitWorktrees(tmp.resolve("wts").toString())
+                .overlayParity(repo.toString(), wt.toString(), List.of(".env", ".envrc"));
+
+        assertEquals("A=1\n", Files.readString(wt.resolve(".env")));
+        assertEquals("export A=1\n", Files.readString(wt.resolve(".envrc")));
+        assertFalse(Files.exists(wt.resolve("not-overlaid.txt")),
+                "overlayParity must copy only the configured candidates, not the whole source tree");
+    }
+
+    /** Criterion 2: both candidates present — the summary line names both and the denominator. */
+    @Test
+    void overlayParityLogsBothCopiedWhenBothCandidatesArePresent(@TempDir Path tmp) throws Exception {
+        reportingLogger.setLevel(Level.INFO);
+        Path repo = initRepo(tmp.resolve("repo"));
+        Files.writeString(repo.resolve(".env"), "A=1\n");
+        Files.writeString(repo.resolve(".envrc"), "export A=1\n");
+        Path wt = bareWorktree(repo, tmp.resolve("wt"), "cb134-both");
+
+        new GitWorktrees(tmp.resolve("wts").toString())
+                .overlayParity(repo.toString(), wt.toString(), List.of(".env", ".envrc"));
+
+        assertTrue(capturedMessages().contains("parity overlay: copied 2 of 2 candidates: .env, .envrc"),
+                "expected the both-copied summary line, got:\n" + capturedMessages());
+    }
+
+    /** Criterion 2: one candidate present, one absent — the summary must name the copied file, the
+     *  denominator, and why the other candidate was not copied. */
+    @Test
+    void overlayParityLogsOneCopiedOneAbsent(@TempDir Path tmp) throws Exception {
+        reportingLogger.setLevel(Level.INFO);
+        Path repo = initRepo(tmp.resolve("repo"));
+        Files.writeString(repo.resolve(".env"), "A=1\n");
+        // .envrc deliberately not created — the absent candidate.
+        Path wt = bareWorktree(repo, tmp.resolve("wt"), "cb134-partial");
+
+        new GitWorktrees(tmp.resolve("wts").toString())
+                .overlayParity(repo.toString(), wt.toString(), List.of(".env", ".envrc"));
+
+        assertTrue(Files.exists(wt.resolve(".env")));
+        assertFalse(Files.exists(wt.resolve(".envrc")));
+        assertTrue(capturedMessages().contains("parity overlay: copied 1 of 2 candidates: .env (.envrc absent)"),
+                "expected the copied/absent summary line, got:\n" + capturedMessages());
+    }
+
+    /** Criterion 3: a tracked candidate is marked --skip-worktree, and that must be named in the log
+     *  with the consequence spelled out — a worker editing it afterward finds git ignoring the
+     *  change, silently, unless this line told it so beforehand. */
+    @Test
+    void overlayParityLogsSkipWorktreeConsequenceForATrackedFile(@TempDir Path tmp) throws Exception {
+        reportingLogger.setLevel(Level.INFO);
+        Path repo = initRepo(tmp.resolve("repo"));
+        Files.writeString(repo.resolve(".env"), "A=1\n");
+        git(repo, "add", ".env");
+        git(repo, "commit", "-q", "-m", "track env");
+        Path wt = bareWorktree(repo, tmp.resolve("wt"), "cb134-tracked");
+        // Change the source after the worktree checkout, so the overlay copy actually overwrites it.
+        Files.writeString(repo.resolve(".env"), "A=2\n");
+
+        new GitWorktrees(tmp.resolve("wts").toString())
+                .overlayParity(repo.toString(), wt.toString(), List.of(".env"));
+
+        assertEquals("A=2\n", Files.readString(wt.resolve(".env")));
+        assertEquals("", status(wt, ".env"),
+                "the skip-worktree'd file must not show as modified even though its content changed");
+        assertTrue(capturedMessages().contains(
+                        "parity overlay marked --skip-worktree (cannot be committed from this worktree): .env"),
+                "expected the skip-worktree consequence line, got:\n" + capturedMessages());
+    }
+
+    /** Criterion 5: null and empty overlay lists return quietly — no exception, no log noise. */
+    @Test
+    void overlayParityWithNoCandidatesLogsNothing(@TempDir Path tmp) throws Exception {
+        reportingLogger.setLevel(Level.INFO);
+        Path repo = initRepo(tmp.resolve("repo"));
+        Path wt = bareWorktree(repo, tmp.resolve("wt"), "cb134-empty");
+        GitWorktrees worktrees = new GitWorktrees(tmp.resolve("wts").toString());
+
+        worktrees.overlayParity(repo.toString(), wt.toString(), null);
+        worktrees.overlayParity(repo.toString(), wt.toString(), List.of());
+
+        assertTrue(reportingAppender.list.isEmpty(),
+                "a null/empty overlay must log nothing, got:\n" + capturedMessages());
+    }
 }

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

@@ -140,7 +140,7 @@ class WorktreeSessionManagerTest {
     void worktreeAcquireRunsParityOverlayWithProfileDefaults() {
         FakeHerdr herdr = new FakeHerdr();
         FakeWorktrees worktrees = new FakeWorktrees().withRepoRoot("/repo").withPrefix("/wt")
-                .track(".envrc")
+                .track(".env")
                 .exists(".claude/settings.local.json");
         SessionManager sessions = new SessionManager(workerService(herdr), worktrees);
 
@@ -151,14 +151,15 @@ class WorktreeSessionManagerTest {
         FakeWorktrees.OverlayCall overlay = worktrees.lastOverlay();
         assertNotNull(overlay);
         assertEquals("/repo", overlay.repoRoot());
-        assertEquals(List.of(".env", ".envrc"),
-                overlay.requested(), "default parity overlay is used when unset");
+        assertEquals(List.of(".env"),
+                overlay.requested(), "default parity overlay is used when unset (CB-148: .env only, "
+                        + ".envrc is no longer defaulted because it is executable shell direnv runs on cd)");
         assertFalse(overlay.requested().contains(".mcp.json"),
                 "CB-525: replicating the primary's MCP config gives a worker the primary's IDE "
                         + "servers, which navigate its edits out of its own worktree");
-        assertEquals(List.of(".envrc"), overlay.copied(),
+        assertEquals(List.of(".env"), overlay.copied(),
                 "existing paths are copied; missing paths are skipped");
-        assertEquals(List.of(".envrc"), overlay.skipWorktree(),
+        assertEquals(List.of(".env"), overlay.skipWorktree(),
                 "tracked copied paths are --skip-worktree'd");
     }