fleetd #391: require reply caller role

The CB-633 allow-list scrub has been dying mid-loop on every fleet01 member
pane and saying nothing. `export UID=` in zsh is not a failed command — it is
a fatal parameter error that terminates the whole sourced file. The blanking
loop is wrapped in `{ ... } 2>/dev/null`, so the message was swallowed and the
report block after the loop never ran.

Root cause found by the fleet01 lead, with xtrace on a live pane's own ZDOTDIR:

    +scrub.zsh:28> _cb633_n=UID
    +scrub.zsh:28> export 'UID='
    +zsh:1> rc=126        <- file aborted

The severity is the INVERSION, and this is their finding, quoted:

  "env lists inherited names first and the names a startup file exports last.
   So the loop blanks the harmless inherited half and dies immediately before
   the operator's own exports — exactly the credentials the policy exists to
   remove. The selection is inverted, not merely partial."

Measured there: UID is name 42 of 57, and a ~/.zshrc decoy at 58 survived on
8 of 8 spawns. "Partial scrub" reads as "we got most of it"; it got precisely
the wrong half.

Fixed with `eval "export ${n}=" 2>/dev/null` rather than a skip-list of the
known-fatal names (UID EUID GID EGID PPID LINENO). A skip-list has to be
complete forever and this is a security control; eval needs no list. Measured:
plain export dies at UID and every later name keeps its value, while the eval
form completes the loop and blanks all of them. PR #396 proposed the skip-list
and is closed in favour of this; its claim that the abort happens "however the
assignment is wrapped" holds for a direct `if ! export` but not for eval,
which reparses in a nested context.

The report now carries `failed N` and `!`-prefixed unblankable names, and
HerdrPeerLauncher WARNs when any name could not be blanked. The old "no report"
WARN no longer claims the daemon knows what the member saw.

Why the suite stayed green: EnvAllowListScrubTest starts zsh from
pb.environment().clear(), and under a cleared parent UID is not an exported
name at all, so the abort could not reproduce in that harness.

Reviewed by mutation, which found a second gap now also closed: the eval is
only safe because names are filtered to ^[A-Za-z_][A-Za-z0-9_]*$. Replacing
that pattern with .* left the class green, so the line the security property
rests on was unpinned. The guard is now re-asserted at the eval site and pinned
by a test. The reachable vector is a VALUE with an embedded newline, not a
hostile name — measured: zsh strips non-identifier env names outright, while
MULTI=$'keep\njunk.fragment' forges 'junk.fragment' as a candidate name out of
its own value.

Closes #394. Refs #396, #388.

CLAUDE.md

@@ -137,7 +137,7 @@ the merge — and merging on a reviewer's word is delegating it by proxy.
 | Answer a member's `fleet_ask` | `fleet_send{turnId, content}` — **not** `sessionId` |
 | Message a **peer lead** on this host | `fleet_send{sessionId: <their terminal>, content}` — `fleet_list` → `leads` reports it. Coordination only, **never** a task |
 | Message a **peer lead** on another daemon or host | `fleet_send{coordId: <their coord-id>, content}` — needs a `coordinator:` block; your own coord-id is in `fleet_list`. Coordination only, **never** a task |
-| Answer a peer lead that messaged you | `fleet_reply{content}` — the one case a lead replies |
+| Answer a peer lead that messaged you | `fleet_send{coordId}` — or `{sessionId}` if they are on this host. **Not** `fleet_reply`: it has no peer route and the publish is refused |
 | Collect a held reply | `fleet_poll{target}` · then `fleet_ack{target, msgId}` |
 | Tear down a member | `fleet_stop{paneId}` |
 
@@ -163,15 +163,21 @@ The traffic between leads is coordination and nothing else:
 3. **Verify a peer exactly as you verify yourself.** Peer status buys nothing: check the claim
    against the code, and re-run the build. A peer's correction gets the same treatment — right or
    wrong on the evidence, not on who said it. Neither of you merges the other's work unreviewed.
+   **N observations are N data points only if they differ in the axis you are trusting.** This cuts
+   both ways. N *failures* blamed on one cause are one data point when the cases share what you are
+   not varying. N *agreeing measurements* are also one data point when they share an instrument —
+   two hosts, two operators and the same formula is one formula, not two confirmations.
 4. **Ask a peer to read your project addendum.** Your addendum is instruction surface: every future
    session on your host obeys it, and a wrong one is obeyed just as faithfully as a right one. The
    author is the worst reader of their own qualifier placement — measured here, one addendum carried
    two defects and a non-author found both. If you have no peer, at least re-read it asking "which
    sentence goes false first, and would a reader reach the caveat before acting?"
 
-Being messaged by a peer does not make you its worker: answer with `fleet_reply`, and push back on
-the substance if it is wrong. A peer that simply complies has thrown away the reason there are two of
-you.
+Being messaged by a peer does not make you its worker: answer the way you would open —
+`fleet_send{coordId}` for another daemon, `fleet_send{sessionId}` on this host — and push back on
+the substance if it is wrong. `fleet_reply` resolves a member's blocked `fleet_send`; a peer's
+coord-id message is durable and non-blocking, so there is nothing for it to resolve. A peer that
+simply complies has thrown away the reason there are two of you.
 
 ### Member (worker or architect) — the turn contract
 

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

@@ -346,7 +346,7 @@ public final class FleetMcp {
                     String self = callerTerminal(exchange);
                     McpSchema.CallToolResult denied = deny(exchange, toolAction("fleet_reply", req.arguments()), self);
                     if (denied != null) return denied;
-                    return reply(messages, self, str(req.arguments(), "content"));
+                    return reply(messages, self, principal(exchange).role(), str(req.arguments(), "content"));
                 };
         // fleet_ask (CB-205): a worker's mid-turn question — identity from the CONNECTION.
         BiFunction<McpSyncServerExchange, McpSchema.CallToolRequest, McpSchema.CallToolResult> askHandler =
@@ -868,19 +868,26 @@ public final class FleetMcp {
     /**
      * {@code fleet_reply}: the worker returns its structured answer, resolving the awaiting send
      * or — when no send is open — queueing the reply in the inbox for later drain (CB-307).
-     * {@code callerTerminal} is resolved from the connection (never an argument); a {@code null}
-     * means the caller is not a known worker (e.g. the primary called it by mistake).
+     * {@code callerTerminal} and {@code callerRole} are resolved from the connection (never an
+     * argument). A {@code null} terminal means the caller is not a known worker. A PRIMARY with a
+     * terminal is a lead and must use {@code fleet_send}, because reply has no peer-lead route.
      *
      * <p>fleetd #365: the result text names which of those actually happened
      * ({@link MessageService.ReplyOutcome#description()}) instead of the single word "delivered"
      * for both — a queued reply is a real success, but it is not the same fact as one that resolved
      * a live waiter, and the caller could not previously tell them apart.
      */
-    static McpSchema.CallToolResult reply(MessageService messages, String callerTerminal, String content) {
+    static McpSchema.CallToolResult reply(MessageService messages, String callerTerminal, Role callerRole, String content) {
         if (callerTerminal == null) {
             return error("fleet_reply is for workers only — could not identify the calling worker "
                     + "from the connection");
         }
+        if (callerRole == Role.PRIMARY) {
+            return error("fleet_reply has no route to a peer lead. Use fleet_send{coordId: ...} for a peer on another "
+                    + "daemon or fleet_send{sessionId: ...} for a peer on this host. fleet_reply resolves a member's "
+                    + "blocked fleet_send, and a peer's coord-id message is durable and non-blocking, so there is "
+                    + "nothing for it to resolve.");
+        }
         // fleetd #302: isBlank, not == null, to match fleet_send's own guard above. MessageService
         // .reply now REJECTS blank content, and this handler is a bare BiFunction with no try/catch
         // around it — so a whitespace-only fleet_reply would leave here as an uncaught

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

@@ -75,9 +75,28 @@ import java.util.stream.Stream;
  * never to the control.
  *
  * <p>The scrub also writes {@code scrub-report.txt} into its own directory: one {@code allowed N of
- * M} line (N = exports left untouched, M = exports present when the scrub ran), then the blanked
- * NAMES — never values. The launcher reads this back at teardown and logs it, because a blocked
- * count next to an unknown denominator is not a finding.
+ * M failed F} line (N = exports left untouched, M = exports present when the scrub ran, F = names
+ * the scrub attempted to blank but could not), then the NAMES — blanked ones bare, unblankable ones
+ * {@code !}-prefixed — never values. The launcher reads this back at teardown and logs it, because a
+ * blocked count next to an unknown denominator is not a finding.
+ *
+ * <p><b>fleetd #394:</b> plain {@code export "$n="} is a FATAL error for a zsh read-only or special
+ * parameter (for example {@code UID}) — it aborts the whole sourced file, so every name still to
+ * come is never blanked and the report above is never written at all. The blanking loop instead
+ * routes each attempt through {@code eval}, which contains that error to the single iteration: the
+ * loop always finishes, and a name that could not be blanked is counted as {@code failed} and
+ * listed {@code !}-prefixed rather than silently disappearing. This is deliberately not a skip-list
+ * of known-bad names — every enumerated name is still attempted, so a name nobody has thought of
+ * yet still gets tried and, if it fails, still gets counted.
+ *
+ * <p>The blanking loop also re-asserts, on its own, the same {@code [A-Za-z_][A-Za-z0-9_]*} shape
+ * check the enumeration loop already applied. Before {@code eval} was introduced a non-conforming
+ * name reaching {@code export "$n="} was harmless either way — the quoting made it inert. With
+ * {@code eval}, the name is spliced into a string and interpreted as shell syntax, so the enumeration
+ * loop's check is no longer sufficient on its own to keep that call site safe — it is a guard on a
+ * different loop, and the two must not silently drift apart. Re-checking right before the
+ * {@code eval} keeps that call site safe by its own reading, independent of whatever the enumeration
+ * loop does or stops doing in a later change.
  */
 public final class EnvAllowListScrub {
 
@@ -132,10 +151,16 @@ public final class EnvAllowListScrub {
     }
 
     /**
-     * A parsed {@code scrub-report.txt}: how many exported variables existed when the scrub ran,
-     * how many were left untouched (allowed), and the NAMES that were blanked. Values never appear.
+     * A parsed {@code scrub-report.txt}: how many exported variables existed when the scrub ran
+     * ({@code total}), how many were left untouched ({@code allowed}), how many the scrub attempted
+     * to blank but could not ({@code failed} — fleetd #394: a zsh read-only/special parameter such
+     * as {@code UID} fatally errors on plain {@code export NAME=}, so those attempts go through
+     * {@code eval} instead so the loop keeps going and the failure is counted rather than left
+     * invisible), and the NAMES in each of the latter two categories. {@code allowed +
+     * blanked.size() + unblankable.size() == total}, and {@code unblankable.size() == failed}.
+     * Values never appear.
      */
-    record ScrubReport(int allowed, int total, List<String> blanked) {
+    record ScrubReport(int allowed, int total, int failed, List<String> blanked, List<String> unblankable) {
     }
 
     /**
@@ -334,38 +359,46 @@ public final class EnvAllowListScrub {
                   _cb633_blank+=("$_cb633_n")
                 done
 
-                # `export UID=` is not a failed command: it is a FATAL zsh parameter error
-                # ("failed to change user ID") that aborts this whole sourced file mid-loop,
-                # leaving every later name unscrubbed and the report below unwritten — silently,
-                # because of the 2>/dev/null. Neither `|| true` nor a `${(t)n}` type guard
-                # contains it; only `eval` does. `eval` is safe here precisely because the loop
-                # above already rejected every name that is not [A-Za-z_][A-Za-z0-9_]*, so
-                # nothing but a bare identifier can reach it.
-                #
-                # Enumerating the special names instead (UID|EUID|GID|EGID|PPID|LINENO) also
-                # works, but only for the ones enumerated: a special that turns up exported on
-                # some other host brings the abort straight back. `eval` contains all of them.
-                #
-                # Then VERIFY. A contained failure is still a failure, so a name that did not
-                # actually blank must not be reported as blanked. It currently falls into the
-                # "allowed" count, which is imprecise in the safe direction; the honest third
-                # count ("tried and could not blank") needs a report-format change and belongs
-                # with fleetd #394, not here.
-                typeset -a _cb633_done
-                _cb633_done=()
+                # fleetd #394: plain `export "$n="` is FATAL for a zsh read-only/special parameter
+                # (e.g. UID) and aborts this whole sourced file — every name still to come is never
+                # blanked, and the report below is never written, silently. `eval` contains that
+                # error to the single iteration instead: it still fails for that one name, but the
+                # loop continues and we can tell allowed / blanked / unblankable apart afterwards.
+                # This is not a skip-list of known-bad names (that would miss the next one nobody
+                # thought of) — every name in _cb633_blank is still attempted, unconditionally.
+                # Every name reaching this loop already passed the identical identifier check in the
+                # enumeration loop above — but that guard is 20 lines away in a different loop, and
+                # this line is about to splice the name into a string handed to `eval`. Before this
+                # fix the name only ever reached `export` quoted ("$n="), which is inert on a
+                # non-identifier string either way; `eval` makes THIS line the only thing standing
+                # between such a string and code execution in the member's pane, so it re-asserts the
+                # same check on its own rather than trusting a guard it does not own. Under normal
+                # operation this can never fire (the enumeration guard already filtered everything
+                # reaching _cb633_blank), so a name caught here is counted as unblankable rather than
+                # silently dropped — it is real evidence that the upstream guard was bypassed.
+                typeset -a _cb633_ok _cb633_unblankable
+                _cb633_ok=()
+                _cb633_unblankable=()
                 for _cb633_n in "${_cb633_blank[@]}"; do
-                  eval "export ${_cb633_n}=" 2>/dev/null
-                  [[ -z "${(P)_cb633_n}" ]] && _cb633_done+=("$_cb633_n")
+                  if [[ ! "$_cb633_n" =~ ^[A-Za-z_][A-Za-z0-9_]*$ ]]; then
+                    _cb633_unblankable+=("$_cb633_n")
+                    continue
+                  fi
+                  if eval "export ${_cb633_n}=" 2>/dev/null; then
+                    _cb633_ok+=("$_cb633_n")
+                  else
+                    _cb633_unblankable+=("$_cb633_n")
+                  fi
                 done
-                _cb633_blank=("${_cb633_done[@]}")
 
                 integer _cb633_kept=$(( _cb633_total - ${#_cb633_blank} ))
                 {
-                  print -r -- "allowed $_cb633_kept of $_cb633_total"
-                  for _cb633_n in "${_cb633_blank[@]}"; do print -r -- "$_cb633_n"; done
+                  print -r -- "allowed $_cb633_kept of $_cb633_total failed ${#_cb633_unblankable}"
+                  for _cb633_n in "${_cb633_ok[@]}"; do print -r -- "$_cb633_n"; done
+                  for _cb633_n in "${_cb633_unblankable[@]}"; do print -r -- "!$_cb633_n"; done
                 } > "$ZDOTDIR/%s" 2>/dev/null
 
-                unset _cb633_done _cb633_allowed _cb633_names _cb633_blank _cb633_n _cb633_total _cb633_kept
+                unset _cb633_allowed _cb633_names _cb633_blank _cb633_ok _cb633_unblankable _cb633_n _cb633_total _cb633_kept
                 """.formatted(names, MemberEnvAllowList.zshCasePattern(), REPORT_FILE);
     }
 
@@ -379,6 +412,11 @@ public final class EnvAllowListScrub {
      * Read and parse {@link #REPORT_FILE} out of a generated ZDOTDIR directory. Returns {@code null}
      * when absent or unreadable (the pane may have been torn down before its login shell ever got to
      * the scrub) — callers treat that as "no measurement available", never as success.
+     *
+     * <p>First line is {@code "allowed <N> of <M> failed <F>"} (fleetd #394 added the trailing
+     * {@code failed <F>} — a count of names the scrub attempted to blank but could not, e.g. a zsh
+     * read-only/special parameter). Every following non-blank line is a name: a bare name was
+     * blanked, a {@code !}-prefixed name was attempted and failed. Values never appear on either.
      */
     static ScrubReport readReport(Path zdotdir) {
         Path report = zdotdir.resolve(REPORT_FILE);
@@ -391,17 +429,24 @@ public final class EnvAllowListScrub {
                 return null;
             }
             String[] parts = lines.getFirst().substring("allowed ".length()).trim().split("\\s+");
-            if (parts.length != 3 || !"of".equals(parts[1])) {
+            if (parts.length != 5 || !"of".equals(parts[1]) || !"failed".equals(parts[3])) {
                 return null;
             }
             List<String> blanked = new ArrayList<>();
+            List<String> unblankable = new ArrayList<>();
             for (int i = 1; i < lines.size(); i++) {
-                if (!lines.get(i).isBlank()) {
-                    blanked.add(lines.get(i));
+                String line = lines.get(i);
+                if (line.isBlank()) {
+                    continue;
+                }
+                if (line.startsWith("!")) {
+                    unblankable.add(line.substring(1));
+                } else {
+                    blanked.add(line);
                 }
             }
             return new ScrubReport(Integer.parseInt(parts[0]), Integer.parseInt(parts[2]),
-                    List.copyOf(blanked));
+                    Integer.parseInt(parts[4]), List.copyOf(blanked), List.copyOf(unblankable));
         } catch (IOException | NumberFormatException e) {
             return null;
         }

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

@@ -1643,11 +1643,19 @@ public abstract class HerdrPeerLauncher implements PeerLauncher {
             log.warn("memberCredentials allow-list: pane {} left no scrub report in {} — the "
                     + "environment scrub cannot be confirmed to have run. Either the pane ended "
                     + "before its shell finished starting, or its shell never read our generated "
-                    + "startup files, in which case that member saw the full host environment.",
+                    + "startup files. Either way, we cannot tell from here whether the scrub ran, "
+                    + "so we do not know what that member's environment contained.",
                     paneId, dir);
         } else {
             log.info("memberCredentials allow-list: pane {} allowed {} of {} environment variables",
                     paneId, report.allowed(), report.total());
+            if (report.failed() > 0) {
+                log.warn("memberCredentials allow-list: pane {} could not blank {} environment "
+                        + "variable(s) — {} (likely a zsh read-only/special parameter) — those "
+                        + "names were left in the member's environment. Confirm none of them is a "
+                        + "credential.",
+                        paneId, report.failed(), report.unblankable());
+            }
             List<String> shaped = report.blanked().stream()
                     .filter(name -> CREDENTIAL_SHAPED_NAME.matcher(name).matches())
                     .toList();

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

@@ -3,6 +3,7 @@ package dev.ltms.fleet.mcp;
 import dev.ltms.fleet.auth.CallerResolver;
 import dev.ltms.fleet.auth.MemberRegistry;
 import dev.ltms.fleet.auth.Principal;
+import dev.ltms.fleet.auth.Role;
 import dev.ltms.fleet.config.FleetConfig;
 import dev.ltms.fleet.guard.SubscriptionGuard;
 import dev.ltms.fleet.herdr.AgentControl;
@@ -28,6 +29,7 @@ import dev.ltms.fleet.placement.BackendQuarantine;
 import dev.ltms.fleet.placement.PlacementPolicies;
 import io.modelcontextprotocol.spec.McpSchema;
 import dev.ltms.fleet.msg.InMemoryReplyInbox;
+import dev.ltms.fleet.msg.ReplyInbox;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 
@@ -77,7 +79,7 @@ class FleetMcpTest {
             Thread.sleep(5);
         }
         assertTrue(rendezvous.isWaiting(target), "send should be accepted for " + target);
-        FleetMcp.reply(messages, target, "received");
+        FleetMcp.reply(messages, target, Role.WORKER, "received");
         assertEquals("received", textOf(send.get(6, TimeUnit.SECONDS)));
     }
 
@@ -110,7 +112,7 @@ class FleetMcpTest {
 
         // fleetd #365: a resolved live send must read distinctly from a merely-queued reply —
         // see replyWithNoPendingSendIsQueuedNotError below for the other case.
-        McpSchema.CallToolResult reply = FleetMcp.reply(messages, "term_a", "LGTM");
+        McpSchema.CallToolResult reply = FleetMcp.reply(messages, "term_a", Role.WORKER, "LGTM");
         assertEquals(MessageService.ReplyOutcome.RESOLVED_SEND.description(), textOf(reply));
 
         McpSchema.CallToolResult res = send.get(6, TimeUnit.SECONDS);
@@ -136,7 +138,7 @@ class FleetMcpTest {
         }
         assertTrue(rendezvous.isWaiting("term_a"), "send should have opened its waiter");
 
-        McpSchema.CallToolResult reply = FleetMcp.reply(messages, "term_a", "async LGTM");
+        McpSchema.CallToolResult reply = FleetMcp.reply(messages, "term_a", Role.WORKER, "async LGTM");
         assertEquals(MessageService.ReplyOutcome.RESOLVED_SEND.description(), textOf(reply));
 
         // Poll until the async send completes and reports the reply.
@@ -183,7 +185,7 @@ class FleetMcpTest {
             Thread.sleep(5);
         }
         assertTrue(rendezvous.isWaiting("term_a"));
-        FleetMcp.reply(messages, "term_a", "done");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "done");
         assertEquals("done", textOf(answer.get(6, TimeUnit.SECONDS)));
 
         McpSchema.CallToolResult done = FleetMcp.poll(messages, ticket, null);
@@ -215,7 +217,7 @@ class FleetMcpTest {
         // fleet_poll{ticket} stuck PENDING forever and later force-failed with a false "session
         // released before it replied" reason. This used to land in the inbox instead (see the old
         // assertion this replaced: messages.drainReplies("term_a").getFirst()...) — that was the bug.
-        FleetMcp.reply(messages, "term_a", "finished after timeout");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "finished after timeout");
         assertEquals("finished after timeout", textOf(FleetMcp.poll(messages, ticket, null)));
         assertTrue(messages.drainReplies("term_a").isEmpty(),
                 "the reply completed its own ticket directly and never touched the inbox");
@@ -269,7 +271,7 @@ class FleetMcpTest {
         }
         assertEquals(MessageService.Phase.FAILED, second.phase());
 
-        FleetMcp.reply(messages, "term_a", "late reply");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "late reply");
         assertEquals("late reply", messages.drainReplies("term_a").getFirst().content());
 
         CompletableFuture<McpSchema.CallToolResult> answer = CompletableFuture.supplyAsync(
@@ -278,7 +280,7 @@ class FleetMcpTest {
         while (!rendezvous.isWaiting("term_a") && System.currentTimeMillis() < deadline) {
             Thread.sleep(5);
         }
-        FleetMcp.reply(messages, "term_a", "done");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "done");
         assertEquals("done", textOf(answer.get(6, TimeUnit.SECONDS)));
     }
 
@@ -331,7 +333,7 @@ class FleetMcpTest {
     void replyWithNoPendingSendIsQueuedNotError() {
         // CB-307: a reply with no open send is now queued in the inbox, not an error.
         // fleetd #365: it must also no longer claim "delivered" — nothing was waiting for it.
-        McpSchema.CallToolResult res = FleetMcp.reply(messages, "term_a", "orphan");
+        McpSchema.CallToolResult res = FleetMcp.reply(messages, "term_a", Role.WORKER, "orphan");
         assertNotEquals(Boolean.TRUE, res.isError(), "a queued reply is not an error");
         assertEquals(MessageService.ReplyOutcome.QUEUED.description(), textOf(res));
 
@@ -341,6 +343,40 @@ class FleetMcpTest {
         assertEquals("orphan", drained.getFirst().content());
     }
 
+    @Test
+    void replyFromLeadIsRefusedBeforeItCanPublishToTheWorkerInbox() {
+        ReplyInbox inboxThatRejectsPublishes = new ReplyInbox() {
+            @Override public void own(String target) { }
+            @Override public void release(String target) { }
+            @Override public void publish(String target, String msgId, String content) {
+                fail("a lead fleet_reply must not publish to the worker inbox");
+            }
+            @Override public List<InboxMessage> peek(String target) { return List.of(); }
+            @Override public void ack(String target, String msgId) { }
+        };
+        MessageService leadMessages = new MessageService(agents, new Injector(agents), new Rendezvous(),
+                inboxThatRejectsPublishes);
+
+        McpSchema.CallToolResult res = assertDoesNotThrow(
+                () -> FleetMcp.reply(leadMessages, "term_lead", Role.PRIMARY, "peer reply"));
+
+        assertTrue(res.isError());
+        assertEquals("fleet_reply has no route to a peer lead. Use fleet_send{coordId: ...} for a peer on another "
+                        + "daemon or fleet_send{sessionId: ...} for a peer on this host. fleet_reply resolves a member's "
+                        + "blocked fleet_send, and a peer's coord-id message is durable and non-blocking, so there is "
+                        + "nothing for it to resolve.",
+                textOf(res));
+    }
+
+    @Test
+    void replyFromUnidentifiedCallerKeepsItsOwnError() {
+        McpSchema.CallToolResult res = FleetMcp.reply(messages, null, Role.PRIMARY, "reply");
+
+        assertTrue(res.isError());
+        assertEquals("fleet_reply is for workers only — could not identify the calling worker from the connection",
+                textOf(res));
+    }
+
     @Test
     void replyWithBlankContentIsACleanToolErrorNotAnUncaughtException() {
         // fleetd #302: MessageService.reply now REJECTS blank content by throwing. fleet_reply's
@@ -351,7 +387,7 @@ class FleetMcpTest {
         // has always used isBlank for exactly this reason.
         for (String blank : new String[] {null, "", "   ", "\n\t"}) {
             McpSchema.CallToolResult res = assertDoesNotThrow(
-                    () -> FleetMcp.reply(messages, "term_a", blank),
+                    () -> FleetMcp.reply(messages, "term_a", Role.WORKER, blank),
                     "blank content must be refused as a tool error, never thrown out of the handler");
             assertEquals(Boolean.TRUE, res.isError(), "blank content is an error result");
             assertTrue(textOf(res).contains("content is required"),
@@ -364,7 +400,7 @@ class FleetMcpTest {
     @Test
     void bridgePollWithTargetDrainsReplies() {
         // A reply with no open send queues it in the inbox.
-        FleetMcp.reply(messages, "term_a", "queued-msg");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "queued-msg");
 
         // fleet_poll with target drains the inbox.
         McpSchema.CallToolResult res = FleetMcp.poll(messages, null, "term_a");
@@ -415,7 +451,7 @@ class FleetMcpTest {
             Thread.sleep(5);
         }
         assertTrue(rendezvous.isWaiting("term_a"), "the answer should have reopened a waiter");
-        McpSchema.CallToolResult reply = FleetMcp.reply(messages, "term_a", "done");
+        McpSchema.CallToolResult reply = FleetMcp.reply(messages, "term_a", Role.WORKER, "done");
         assertEquals(MessageService.ReplyOutcome.RESOLVED_SEND.description(), textOf(reply));
         assertEquals("done", textOf(answer.get(6, TimeUnit.SECONDS)));
     }
@@ -1254,13 +1290,13 @@ class FleetMcpTest {
     @Test
     void bridgeAckRemovesSpecificReply() {
         // Queue a reply and capture its msgId.
-        FleetMcp.reply(messages, "term_a", "orphan");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "orphan");
         var before = messages.drainReplies("term_a");
         assertEquals(1, before.size(), "one reply in the inbox");
         String msgId = before.getFirst().msgId();
 
         // Publish the same reply again and ack it via fleet_ack surface.
-        FleetMcp.reply(messages, "term_a", "orphan-again");
+        FleetMcp.reply(messages, "term_a", Role.WORKER, "orphan-again");
         var peeked = messages.drainReplies("term_a");
         assertEquals(1, peeked.size(), "one fresh reply in the inbox");
 

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

@@ -120,40 +120,128 @@ class EnvAllowListScrubTest {
     }
 
     /**
-     * A pane inherits {@code UID}; a cleared test parent does not. The scrub must survive it.
+     * fleetd #394: the actual defect. Plain {@code export "$n="} is FATAL for a zsh read-only or
+     * special parameter (e.g. {@code UID}) and aborts the whole sourced file — every name still to
+     * come is never blanked, and the {@code scrub-report.txt} below is never written at all,
+     * silently ({@code 2>/dev/null} swallows the error). This plants an unblankable, exported,
+     * read-only variable in the MIDDLE of the names the scrub attempts to blank, with two more
+     * names after it, and asserts that both of those later names are STILL blanked and the report
+     * is STILL written with the failure counted — a test that only checked names BEFORE the failure
+     * point would pass today and prove nothing.
      *
-     * <p>Every other test here starts zsh from a CLEARED environment, so {@code UID} is never an
-     * exported name and never reaches the blanking loop. In a real member pane it is exported and
-     * it IS reached — and {@code export UID=} is a fatal zsh parameter error that aborts the whole
-     * sourced file, leaving every later name unscrubbed and writing no report at all. The abort is
-     * silent: the loop is wrapped in {@code 2>/dev/null}.
+     * <p>The planted name is a made-up one ({@code FLEETD_TEST_UNBLANKABLE}), not {@code UID} or
+     * any other name a skip-list might already know about — invariant 1 is that the loop survives
+     * ANY unblankable name, not a known one, so the test must not lean on one either.
      *
-     * <p>The assertion is deliberately "a report exists" rather than "the canary is blanked". The
-     * report is written by the last statement in the file, so its presence proves the script ran
-     * to completion; the canary alone would depend on where it happens to sit in {@code env} order.
-     * Both are checked, but only the first one fails deterministically without the fix.
+     * <p>Exercises the real artefact: {@link EnvAllowListScrub#scrubScript} is run verbatim under a
+     * real {@code /bin/zsh}, not just asserted on as a Java string. The four planted names are
+     * exported one at a time via {@code typeset -x}/{@code typeset -rx} immediately before the
+     * script runs, in a fixed order — zsh's {@code export}/{@code typeset -x} appends to the
+     * process's environment table in call order (verified empirically: a freshly-exported name
+     * always sorts after every inherited one and after every earlier freshly-exported name in
+     * {@code command env}'s own output), which is what makes the "middle" position deterministic
+     * here, unlike relying on the OS's own inherited-environment order.
      */
     @Test
-    void scrubSurvivesAnInheritedUidTheWayARealPaneHasIt(@TempDir Path tmp) throws Exception {
+    void unblankableNameInTheMiddleDoesNotAbortNamesAfterIt(@TempDir Path tmp) throws Exception {
         assumeTrue(Files.isExecutable(ZSH), "/bin/zsh not present — nothing to prove here");
-        Set<String> allowed = MemberEnvAllowList.derive(List.of());
-        Path zdotdir = EnvAllowListScrub.generate(tmp, allowed);
 
-        // The production shape: UID present and exported, as every pane shell inherits it.
-        Map<String, String> paneLikeParent = Map.of(
-                "HOME", System.getProperty("user.home"),
-                "PATH", "/usr/bin:/bin",
-                "SHELL", "/bin/zsh",
-                "UID", "1000",
-                "CB633_CANARY", "must-not-survive-the-scrub");
-        Set<String> survivors = exportedNamesFromCleanParent(paneLikeParent, zdotdir);
+        // Only ZDOTDIR is allowed — it must survive the scrub itself, since the report is written
+        // to "$ZDOTDIR/..." AFTER the blanking loop runs; if ZDOTDIR were blanked as a side effect,
+        // the report write would silently go to the wrong place instead of testing anything.
+        String script = EnvAllowListScrub.scrubScript(Set.of("ZDOTDIR"));
+        String setup = """
+                typeset -x FLEETD_TEST_BEFORE=1
+                typeset -rx FLEETD_TEST_UNBLANKABLE=1
+                typeset -x FLEETD_TEST_AFTER_A=1
+                typeset -x FLEETD_TEST_AFTER_B=1
+                """;
 
-        EnvAllowListScrub.ScrubReport report = EnvAllowListScrub.readReport(zdotdir);
-        assertNotNull(report,
-                "an inherited UID must not abort the scrub — no report means the file died mid-loop "
-                        + "and every name after UID in `env` order was left unscrubbed");
-        assertFalse(survivors.contains("CB633_CANARY"),
-                "a non-allow-listed name must still be blanked when UID is in the environment");
+        ProcessBuilder pb = new ProcessBuilder("/bin/zsh");
+        pb.environment().clear();
+        pb.environment().put("PATH", "/usr/bin:/bin");
+        pb.environment().put("ZDOTDIR", tmp.toAbsolutePath().toString());
+        pb.redirectError(ProcessBuilder.Redirect.DISCARD);
+        Process zsh = pb.start();
+        zsh.getOutputStream().write((setup + script).getBytes(StandardCharsets.UTF_8));
+        zsh.getOutputStream().flush();
+        zsh.getOutputStream().close();
+        assertTrue(zsh.waitFor(60, java.util.concurrent.TimeUnit.SECONDS),
+                "the scrub script did not exit within 60s");
+        assertEquals(0, zsh.exitValue(),
+                "the scrub script itself must never abort — an unblankable name must not kill the "
+                        + "sourced file");
+
+        EnvAllowListScrub.ScrubReport report = EnvAllowListScrub.readReport(tmp);
+        assertNotNull(report, "the report must still be written even though one name could not be "
+                + "blanked — a report that silently never appears is the #394 bug");
+        assertTrue(report.blanked().contains("FLEETD_TEST_BEFORE"),
+                "sanity: the name before the unblankable one must be blanked");
+        assertTrue(report.blanked().contains("FLEETD_TEST_AFTER_A"),
+                "the FIRST name AFTER the unblankable one must still be blanked — before the fix, "
+                        + "the whole loop aborted at the unblankable name and every later name was "
+                        + "silently left untouched");
+        assertTrue(report.blanked().contains("FLEETD_TEST_AFTER_B"),
+                "the SECOND name after the unblankable one must also still be blanked");
+        assertEquals(1, report.failed(),
+                "exactly one attempted name could not be blanked, and that count must be visible "
+                        + "without reading the member's environment");
+        assertEquals(List.of("FLEETD_TEST_UNBLANKABLE"), report.unblankable(),
+                "the unblankable name is reported by name, not silently dropped");
+    }
+
+    /**
+     * fleetd #394 follow-up: the blanking loop's {@code eval "export ${n}="} splices {@code n} into
+     * a string that zsh then interprets as shell syntax. That is only safe because every name
+     * reaching {@code _cb633_blank} already passed an identifier check in the ENUMERATION loop
+     * (20 lines away, in a different loop) — so the fix re-asserts the identical check immediately
+     * before the {@code eval} call, rather than trusting that distant guard to keep holding.
+     *
+     * <p>This test plants a value with an embedded newline, exploiting the exact "junk from
+     * multi-line values" gap the enumeration loop's own comment already documents: {@code command
+     * env}'s text output is read line-by-line, so a value's second line becomes a spurious extra
+     * "name" that was never a real exported variable. The fragment used here ({@code
+     * junk.fragment}) is merely non-conforming (it contains a dot) — never command-shaped; this
+     * test must never demonstrate command execution and plants no command-shaped payload.
+     *
+     * <p>Exercises the real artefact end-to-end: {@link EnvAllowListScrub#scrubScript} runs
+     * verbatim under a real {@code /bin/zsh}, exactly as {@code generate()} would produce it — this
+     * is not a synthetic call into just the blanking loop.
+     */
+    @Test
+    void nonIdentifierJunkFromAMultilineValueIsSkippedNotBlankedOrUnblankable(@TempDir Path tmp)
+            throws Exception {
+        assumeTrue(Files.isExecutable(ZSH), "/bin/zsh not present — nothing to prove here");
+
+        String script = EnvAllowListScrub.scrubScript(Set.of("ZDOTDIR"));
+
+        ProcessBuilder pb = new ProcessBuilder("/bin/zsh");
+        pb.environment().clear();
+        pb.environment().put("PATH", "/usr/bin:/bin");
+        pb.environment().put("ZDOTDIR", tmp.toAbsolutePath().toString());
+        // Embedded newline: `command env`'s own text output splits this into two lines, and the
+        // second ("junk.fragment") has no "=" at all, so `cut -d= -f1` returns it unchanged as a
+        // spurious candidate "name" — it was never an actual exported variable by that name.
+        pb.environment().put("FLEETD_TEST_MULTILINE", "keep\njunk.fragment");
+        pb.redirectError(ProcessBuilder.Redirect.DISCARD);
+        Process zsh = pb.start();
+        zsh.getOutputStream().write(script.getBytes(StandardCharsets.UTF_8));
+        zsh.getOutputStream().flush();
+        zsh.getOutputStream().close();
+        assertTrue(zsh.waitFor(60, java.util.concurrent.TimeUnit.SECONDS),
+                "the scrub script did not exit within 60s");
+        assertEquals(0, zsh.exitValue(), "the scrub script must reach its end");
+
+        EnvAllowListScrub.ScrubReport report = EnvAllowListScrub.readReport(tmp);
+        assertNotNull(report, "the report must still be written");
+        assertTrue(report.blanked().contains("FLEETD_TEST_MULTILINE"),
+                "sanity: the real, identifier-shaped variable must still be blanked normally");
+        assertFalse(report.blanked().contains("junk.fragment"),
+                "a non-identifier fragment is not a real variable and must never be blanked");
+        assertFalse(report.unblankable().contains("junk.fragment"),
+                "a non-identifier fragment must never even become a candidate the blanking loop "
+                        + "attempts — it must be filtered before either guard has to catch it, so "
+                        + "it is neither blanked nor counted as a failed attempt");
     }
 
     /** A group-shared ZDOTDIR still lets the member truncate and write its pre-created receipt. */