fleetd#323: close the reload-classifier drift with a reflection coverage test

ConfigRef.sameLaunchSettings' javadoc claimed it compares every component
the launcher reads at spawn. It missed ideProjectDir, ideOpenCommand and
autoCompactWindow, and changedDeferredKeys separately missed worktreeGroup
(baked into the same GitWorktrees as worktreeRoot, Fleetd.java:251). A
reload that changed only one of those keys reported "config reloaded" with
nothing deferred, and the running daemon kept the old value.

Fix the four instances, and add ConfigRefProfileCoverageTest: it enumerates
every FleetConfig.Profile record component by reflection, mutates each one
not in the new ConfigRef.LAUNCH_SETTINGS_EXCLUDED set on a base profile,
and asserts sameLaunchSettings actually notices — so a fifth missed field
fails the build by name instead of drifting silently. It also prints its
own denominator (26 components, 23 compared, 3 excluded) per the ticket's
requirement that a checker must be able to state what it checked.

Also add the `profile` field itself to the comparison (it was neither
compared nor excluded before this fix — the coverage test surfaced it).

Rewrote the sameLaunchSettings javadoc to describe what the coverage test
actually guarantees instead of repeating the unchecked claim.

.claude/skills/hunter/SKILL.md

@@ -0,0 +1,102 @@
+---
+name: hunter
+description: Defect-hunt procedure for a fleetd worker — sweep an assigned package for real bugs and report several ranked findings without fixing anything. Load this when the lead asks you to hunt or audit a scope rather than review one diff. Do NOT load `reviewer` for this; the two want different output.
+---
+
+# Hunter worker — procedure
+
+The turn contract (one `fleet_reply`, `fleet_ask` for the lead's decisions, honest reporting,
+never merge) is in **`CLAUDE.md` → Bridge communication → Worker** and already applies.
+
+**This skill is not `reviewer`.** `reviewer` judges one diff and reports the *single* most
+important issue in about 90 words. A hunt sweeps a whole package and reports *several* findings
+in a long structured form. Loading both gives you two contradictory output contracts, and the
+usual result is a worker that writes a good report into its terminal and ends the turn without
+sending it. Load exactly one.
+
+## 0. Read this before you read code: how the report gets home
+
+Your terminal reaches nobody. The lead sees **only** the text inside your `fleet_reply` call.
+
+A long report is exactly the case where this goes wrong, so plan for it:
+
+- **Write the report into the `fleet_reply` argument itself.** Do not compose it in your terminal
+  and then summarise it into the call.
+- If the report is long, **send it anyway** — one `fleet_reply` with everything.
+- If you end the turn without replying, the bridge scrapes your pane instead. That scrape carries
+  at most the last 4000 characters, and on a hunt it usually captures the tail of the lead's own
+  brief rather than your findings. The lead then has nothing and has to ask you again.
+
+## 1. Change nothing
+
+A hunt is read-only. Do not edit a production file, do not "quickly fix" what you find, and do
+not run a formatter. You may run the build and tests to *check* a claim, and you should say so
+when you did.
+
+## 2. Read the whole scope first
+
+Read every file in the assigned package before you judge any of it. A defect that a caller
+elsewhere in the same package makes unreachable is not a defect, and you cannot know that from
+one file.
+
+Stay inside the scope. If a defect there depends on a class outside it, read that class to
+confirm — but the defect itself must live in the scope you were given.
+
+## 3. The bar — this matters more than the count
+
+**Name the path into the bad state.** Say which caller, in which state, reaches it. A defect on
+paper is not a reachable defect. If you cannot name that path, keep the finding but mark it
+`unproven` and say exactly what you could not check. Do not drop it, and do not dress it up.
+
+**Say which direction the harm goes.** Data loss, privilege escalation and silent wrong answers
+are worth reporting even when the window is narrow. A finding whose worst outcome is a worse log
+line is not worth a block.
+
+Two workers once ran the same scope: the one that applied the direction-of-harm filter found ten
+real defects, the one that did not found none. Fewer findings the lead can act on beat many the
+lead has to triage.
+
+## 4. Shapes that have produced real merged fixes here
+
+Read for these first:
+
+1. **A one-way gate.** A guard added after an incident closes only the direction that incident
+   came from. Do not only ask what closes the gate — ask **which states still open it**.
+2. **A value read once, then used later to authorise something destructive**, after something
+   else has had a chance to change it.
+3. **A failure downgraded to a value that looks like a legitimate result** — `-1`, `null`, an
+   empty list, `false` — which a caller then trusts.
+4. **A lock held for one half of a read-modify-write and not the other**, or two collections
+   updated under different locks.
+5. **A comment or javadoc stating an invariant the code no longer keeps.** Comments are
+   load-bearing in this repo; a stale one has already caused a bug.
+
+## 5. What you cannot check, and must not claim you did
+
+- `fleetd/fleetd.yaml` is gitignored and **absent from your worktree**. You cannot read it. If a
+  finding depends on live configuration, name the key and say you could not check it.
+- `.mcp.json`, `opencode.json` and `.autoenv` in your worktree are neutralised stubs, not the
+  repo's real files.
+- The `wiki/` submodule pointer is months old. Do not cite it.
+
+Reporting a fact you took from the lead's brief as something you measured yourself is a false
+report, even when the fact is correct. Say where each fact came from.
+
+## 6. The report — what goes in `fleet_reply`
+
+One block per finding, most severe first:
+
+```
+FINDING N — <one line>
+file:line
+Path in:        <which caller, in which state, reaches this>
+Direction:      <data loss | escalation | silent wrong answer | outage | ...>
+Window/trigger: <when it actually happens>
+Confidence:     <confirmed by reading | unproven — say what you could not check>
+Why nothing else catches it: <the guard or test you checked, and why it misses>
+```
+
+End with one line naming every file you read, so the lead knows the denominator.
+
+**Nothing clears the bar?** Reply `NO FINDINGS`, name the files you read, and say what you ruled
+out. A clean sweep is a valid result; an invented defect is worse than none.

.claude/skills/reviewer/SKILL.md

@@ -10,6 +10,11 @@ never merge) is in **`CLAUDE.md` → Bridge communication → Worker** and alrea
 skill is only the *review procedure*: how to work the scope, and the exact shape of what you
 send back.
 
+**Wrong skill for a sweep.** This one reviews *one* diff or scope and reports the *single* most
+important issue. If the lead asked you to hunt or audit a whole package for several defects, load
+`hunter` instead and ignore this file — the two want different output, and following both is how a
+worker ends its turn with a good report that never gets sent.
+
 ## 1. Read the whole scope before you judge
 
 The delegation names your scope — a file, a diff, a PR, a function. **Read all of it first.**

CLAUDE.md

@@ -200,8 +200,13 @@ must obey belongs in the charter, not here.
   adapter, with a message naming the credential and the remaining seconds ("cooling off after
   repeated backend errors") — distinct wording from a quarantine refusal, so don't conflate the
   two when reading a spawn failure.
-- **Skills available to delegate:** `implementer` (worktree → commit → push → own PR) and
-  `reviewer` (scoped review → one structured finding). Name one in every delegation.
+- **Skills available to delegate:** `implementer` (worktree → commit → push → own PR),
+  `reviewer` (one diff → one structured finding) and `hunter` (sweep a package → several ranked
+  findings, change nothing). Name exactly one in every delegation. **`reviewer` and `hunter` are
+  not interchangeable** — `reviewer` caps the answer at one finding in about 90 words, so naming
+  it for a multi-finding sweep hands the worker two contradictory output contracts. That has
+  already cost three workers' turns: each wrote a good report to its terminal and ended the turn
+  with no `fleet_reply`, and the scrape returned the tail of the brief instead.
 - **Primary-side skills** (not delegation playbooks — a worker cannot use them):
   `port-to-opencode` (make an OpenCode session a participant in this workspace) and
   `fleets-status` (report every fleet that shares one LavinMQ instance).

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

@@ -236,7 +236,15 @@ public final class CallerResolver {
         // loopback-trust: same-host callers that are not workers are the primary. A non-loopback
         // caller is anonymous even here — and startup refuses that combination anyway
         // (FleetConfig.validateAuthExposure), so this is defence in depth, not the control.
-        return isLoopback(remoteAddr) ? Principal.primary(c.pid()) : Principal.anonymous();
+        //
+        // fleetd #317: "not a worker" must not be conflated with "identity unresolved". The real
+        // primary is a real process — its pid resolves (c.resolved()), it just owns no herdr pane.
+        // A caller whose peer-PID lookup failed (LsofPeerPidLookup's -1 sentinel — on any failure,
+        // silently including "lsof found no match") has no such pid, and PaneLocator's own javadoc
+        // already names what happens if that case is handed the primary role: a worker→primary
+        // escalation. So an unresolved caller is refused (ANONYMOUS — the same clean, already-tested
+        // "authenticated as nothing" outcome used everywhere else in this method), never promoted.
+        return isLoopback(remoteAddr) && c.resolved() ? Principal.primary(c.pid()) : Principal.anonymous();
     }
 
     private boolean presentedTokenMatches(String authorizationHeader) {

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

@@ -39,12 +39,18 @@ import java.util.function.Supplier;
  *       keeps the old value until a restart: {@code lifecycle:}, {@code leadHeartbeat:},
  *       {@code spawnReadyTimeoutMs} / {@code spawnReadyPollMs}, {@code quarantineCooldownSeconds}
  *       (CB-578 stage B — baked once into the {@code BackendQuarantine} built at startup),
- *       {@code guard:}, {@code worktreeRoot:}, adding or removing a profile (a new backend needs its own launcher,
+ *       {@code guard:}, {@code worktreeRoot:} and {@code worktreeGroup:} (both baked once into the
+ *       {@code GitWorktrees} built at {@code Fleetd.java:251} and never rebuilt — fleetd #323
+ *       instance 2 found {@code worktreeGroup} missing from this list and from
+ *       {@link #changedDeferredKeys}), adding or removing a profile (a new backend needs its own launcher,
  *       which is constructed once), <em>and an existing profile's launch settings</em> —
  *       {@code model}, {@code baseUrl}, {@code argv}, {@code env}, {@code mcpUrl},
  *       {@code exhaustedPattern} (CB-578 stage A — compiled once into {@code Fleetd.main}'s
  *       pattern map at startup), {@code errorPattern} (fleetd #201 Unit 5 — compiled once into
- *       {@code Fleetd.main}'s backend-error pattern map at startup, the same way), and the rest.
+ *       {@code Fleetd.main}'s backend-error pattern map at startup, the same way),
+ *       {@code ideProjectDir} / {@code ideOpenCommand} / {@code autoCompactWindow} (fleetd #323
+ *       instance 1 — all three are read at spawn off the same frozen profile map and were missing
+ *       from {@link #sameLaunchSettings}), and the rest of {@link #sameLaunchSettings}.
  *       {@code credentialId} (CB-578 stage B) is NOT on
  *       this list — it is read live off the config supplier at every quarantine check and
  *       exhaustion event, exactly like {@code weight} / {@code maxLoad}, so it is hot instead.
@@ -221,6 +227,13 @@ public final class ConfigRef implements Supplier<FleetConfig> {
         if (!Objects.equals(old.worktreeRoot(), fresh.worktreeRoot())) {
             changed.add("worktreeRoot");
         }
+        // Baked into the same GitWorktrees as worktreeRoot (Fleetd.java:251) and never rebuilt
+        // either — see the class doc. Missing this check was fleetd #323 instance 2: a reload
+        // that changed only worktreeGroup reported "config reloaded" with nothing deferred, and
+        // newly provisioned worktrees kept the old sharing behaviour.
+        if (!Objects.equals(old.worktreeGroup(), fresh.worktreeGroup())) {
+            changed.add("worktreeGroup");
+        }
         if (!Objects.equals(old.spawnReadyTimeoutMs(), fresh.spawnReadyTimeoutMs())
                 || !Objects.equals(old.spawnReadyPollMs(), fresh.spawnReadyPollMs())) {
             changed.add("spawnReady*");
@@ -265,14 +278,34 @@ public final class ConfigRef implements Supplier<FleetConfig> {
     }
 
     /**
-     * Whether two versions of a profile would launch a peer identically. Compares every component
-     * the launcher reads at spawn; {@code weight}, {@code maxLoad} and {@code credentialId} are
-     * excluded because those are read live (by the placement policy and, for credentialId, by
-     * {@code CompositePeerLauncher}/the CB-578 stage B exhaustion sink) and really do take effect on
-     * the next spawn.
+     * {@link FleetConfig.Profile} record components deliberately left out of
+     * {@link #sameLaunchSettings} because they are read <em>live</em>, not baked in at spawn — see
+     * the class doc's <em>Hot</em> bullet. {@code weight} and {@code maxLoad} are read live by the
+     * placement policy on every spawn; {@code credentialId} is read live by
+     * {@code CompositePeerLauncher} and the CB-578 stage B exhaustion sink. Nothing else is
+     * excluded — see {@code sameLaunchSettingsComparesEveryProfileComponentOrExcludesIt} in
+     * {@code ConfigRefProfileCoverageTest}, which enumerates every {@code Profile} record component
+     * by reflection and fails the build if one is neither compared below nor named here.
      */
-    private static boolean sameLaunchSettings(FleetConfig.Profile a, FleetConfig.Profile b) {
-        return Objects.equals(a.baseUrl(), b.baseUrl())
+    static final Set<String> LAUNCH_SETTINGS_EXCLUDED = Set.of("weight", "maxLoad", "credentialId");
+
+    /**
+     * Whether two versions of a profile would launch a peer identically.
+     *
+     * <p>This must compare every {@link FleetConfig.Profile} record component except the three in
+     * {@link #LAUNCH_SETTINGS_EXCLUDED}. That is not a claim this javadoc can make good on by
+     * itself — a javadoc saying "compares every component" is exactly what fleetd #323 found to be
+     * false for three fields (and a sibling method's field list, for a fourth). The actual
+     * guarantee comes from {@code ConfigRefProfileCoverageTest}: it enumerates every record
+     * component of {@code FleetConfig.Profile} by reflection, mutates each one not in
+     * {@code LAUNCH_SETTINGS_EXCLUDED} on a base profile, and asserts this method reports a
+     * difference — so a new component that is neither compared here nor added to
+     * {@code LAUNCH_SETTINGS_EXCLUDED} (with a reason) fails that test by name, rather than
+     * silently reporting "config reloaded" for a value the daemon never picked up.
+     */
+    static boolean sameLaunchSettings(FleetConfig.Profile a, FleetConfig.Profile b) {
+        return Objects.equals(a.profile(), b.profile())
+                && Objects.equals(a.baseUrl(), b.baseUrl())
                 && Objects.equals(a.model(), b.model())
                 && Objects.equals(a.configDir(), b.configDir())
                 && Objects.equals(a.tokenEnv(), b.tokenEnv())
@@ -298,6 +331,16 @@ public final class ConfigRef implements Supplier<FleetConfig> {
                 // fleetd #201 Unit 5: errorPattern is compiled once into Fleetd.main's backend-error
                 // pattern map at startup (see BackendErrorPatternLookup wiring), the same way
                 // exhaustedPattern is — a reload never re-reads it either.
-                && Objects.equals(a.errorPattern(), b.errorPattern());
+                && Objects.equals(a.errorPattern(), b.errorPattern())
+                // fleetd #323 instance 1: ideProjectDir and ideOpenCommand are read at spawn off the
+                // same frozen profile map as ideMcpUrl above (ClaudeCodeLauncher.java:267/269,
+                // OpenCodeLauncher.java:474/480/486) and were missing from this comparison.
+                && Objects.equals(a.ideProjectDir(), b.ideProjectDir())
+                && Objects.equals(a.ideOpenCommand(), b.ideOpenCommand())
+                // fleetd #323 instance 1: autoCompactWindow is read at spawn the same way
+                // (ClaudeCodeLauncher.java:926, OpenCodeLauncher.java:650). Comparing it here only
+                // makes the reload REPORT that a restart is needed — it deliberately does not make
+                // autoCompactWindow take effect live, which is a separate, larger change.
+                && Objects.equals(a.autoCompactWindow(), b.autoCompactWindow());
     }
 }

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

@@ -36,6 +36,25 @@ public final class ConnectionIdentity {
      * primary / an off-host client) and its {@code pid} (or {@code -1} if not resolvable).
      */
     public record Caller(String terminal, long pid) {
+
+        /**
+         * Whether the OS peer-PID lookup actually succeeded — {@code false} means {@code pid} is
+         * the {@code -1} sentinel, not a real process id, so this caller's identity could not be
+         * established at all. That is a different fact from a real pid that simply owns no worker
+         * pane (the primary's own connection): the primary is {@code resolved()} and has a
+         * {@code null terminal}; an unresolvable caller is {@code !resolved()} and also has a
+         * {@code null terminal}. The two look identical through {@link #terminal} alone, which is
+         * exactly how fleetd #317 happened — a failed {@code lsof} lookup and a genuine primary both
+         * fell through to {@code Principal.primary(...)}.
+         *
+         * <p>Centralised here, next to the sentinel it tests, for the same reason
+         * {@link ConnectionIdentity#isLoopback} is centralised rather than left for each caller to
+         * reimplement: a raw {@code pid > 0} check duplicated at every call site is precisely the
+         * "one rule, two copies" shape that let #305 drift.
+         */
+        public boolean resolved() {
+            return pid > 0;
+        }
     }
 
     /** Resolve the caller's terminal and PID from one peer-PID lookup. */
@@ -71,13 +90,27 @@ public final class ConnectionIdentity {
      * {@code 127.0.0.1:8765} with a source address of {@code 127.0.0.2} — measured on the Linux
      * fleet host, where binding that source succeeds.
      *
-     * <p><strong>Being strict here does not make the daemon safer; it makes it unsafe.</strong>
-     * That reads backwards, so it is worth stating plainly. This predicate does not decide whether
-     * a caller is trusted — it decides whether the caller's identity is <em>resolved at all</em>.
-     * Returning false means {@link #resolve} answers "no terminal", and downstream a caller with no
-     * terminal is treated as the primary under loopback-trust. So every address excluded here is an
-     * address on which a worker silently becomes the lead. Widening a check normally weakens it;
-     * widening this one is what closes the hole.
+     * <p><strong>What excluding an address costs, stated as it is today.</strong> This paragraph
+     * used to say that narrowing this range turned a worker into the lead, and that widening the
+     * check was what closed the hole. That was true only while there were <em>two</em> definitions
+     * that disagreed: {@code ConnectionIdentity} skipped the identity lookup for {@code 127.0.0.2}
+     * while {@code CallerResolver} read the same address as loopback and granted the primary role.
+     * #305 removed the second copy, and with one shared definition the old sentence no longer holds.
+     *
+     * <p>Measured on 2026-09-04 by narrowing this method back to exactly {@code 127.0.0.1} and
+     * running {@code CallerResolverTest} and {@code ConnectionIdentityTest}: a caller from
+     * {@code 127.0.0.2} then resolves to {@code ANONYMOUS}, not {@code PRIMARY} — for a worker
+     * ({@code aWorkerOnAnyLoopbackSourceAddressIsStillAWorkerNotThePrimary}) and for a non-worker
+     * ({@code aNonWorkerOnAnyLoopbackSourceAddressIsStillThePrimary}) alike. Excluding an address
+     * now <em>refuses</em> its caller; it does not promote one.
+     *
+     * <p>So keep the whole range, but for the plain reason: a genuine worker or primary that
+     * connects from {@code 127.0.0.2} must be identifiable at all, and narrowing this predicate
+     * locks it out. That is an outage, and an outage is the direction to fail in — which is exactly
+     * why the range must not be narrowed casually and also why doing so is no longer a security
+     * hole. This predicate still does not decide whether a caller is trusted; it decides whether the
+     * caller's identity is <em>resolved at all</em>. What makes an unresolved caller safe is
+     * {@link Caller#resolved()} (#317), not this method.
      */
     public static boolean isLoopback(String addr) {
         if (addr == null) {

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

@@ -41,6 +41,14 @@ public final class LsofPeerPidLookup implements PeerPidLookup {
             if (!p.waitFor(2, TimeUnit.SECONDS)) {
                 p.destroyForcibly();
             }
+            if (found < 0) {
+                // fleetd #317: this is the silent path — lsof ran clean and simply reported no
+                // matching process (e.g. queried before the OS socket table settles). Previously
+                // this logged nothing at all, which is exactly why the escalation went unnoticed;
+                // the exception path below already logs. A caller now refused because of this is
+                // still refused (never promoted) — this line only makes the refusal diagnosable.
+                log.debug("lsof peer-pid lookup for port {} found no matching process", port);
+            }
             return found;
         } catch (Exception e) {
             log.debug("lsof peer-pid lookup for port {} failed: {}", port, e.getMessage());

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

@@ -385,9 +385,12 @@ public final class CompositePeerLauncher implements PeerLauncher {
             }
         }
 
+        // unreachable.size() counts DISTINCT profiles, not attempts (a HashSet dedupes a profile
+        // added twice) — say "distinct" so the count matches the sentence and the profile list that
+        // follows, rather than reading as a count of attempts made (fleetd #315).
         throw new PeerUnreachableException(
                 "no reachable worker profile available after trying " + unreachable.size()
-                        + " candidate(s): " + String.join(", ", unreachable));
+                        + " distinct candidate(s): " + String.join(", ", unreachable));
     }
 
     /**

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

@@ -187,6 +187,24 @@ public final class MessageService {
         private volatile Long completedNanos;
         private volatile Reply question;
         private volatile String turnId;
+        /**
+         * Set when this task's {@code fleet_ask} lapsed with no answer (fleetd #307):
+         * {@link #clearAsyncQuestion} then forgets {@link #turnId} (nulls it and drops the task from
+         * {@code asyncTasksByTurn}) so {@link #hasAsyncQuestion} stops reporting the target BUSY — a
+         * later {@code fleet_send} to it must be accepted, not refused. But the worker's turn is
+         * still genuinely live: it resumed on its own and will eventually call its real
+         * {@code fleet_reply}. Losing {@link #turnId} loses {@link #askAnsweredAsyncTasks}' only
+         * signal that such a reply belongs to this task, so that reply used to fall straight to the
+         * inbox and strand — {@code fleet_poll} stayed {@code PENDING} forever, later force-failed by
+         * {@link #abandon} with the misleading "session released before it replied". This flag is a
+         * second, independent signal that survives the forgetting: {@link #askAnsweredAsyncTasks}
+         * accepts it in place of a live {@link #turnId}, without ever re-adding the task to
+         * {@code asyncTasksByTurn} (so the BUSY release is untouched). Cleared implicitly once
+         * {@link #future} resolves — every match in {@link #askAnsweredAsyncTasks} already requires
+         * {@code !future.isDone()}, so a task that recovered (or was later failed by
+         * {@link #abandon}) can never match again regardless of this flag's value.
+         */
+        private volatile boolean askTimedOut;
 
         private Task(String ticket, String target, LongSupplier nowNanos) {
             this.ticket = ticket;
@@ -395,18 +413,16 @@ public final class MessageService {
      * {@link Rendezvous#resolveQuestion} must keep today's {@code NO_WAITER} behaviour — questions
      * are interactive and must never be queued.
      *
-     * <p><strong>Ambiguous match also falls to the inbox.</strong> {@link #askAnsweredAsyncTasks}
-     * cannot actually return more than one entry today (see its own javadoc for why — in short,
-     * {@link #hasAsyncQuestion} keeps a target BUSY, so no second task can reach this state, for as
-     * long as an earlier one's {@code turnId} is still stamped). That is an emergent guarantee from
-     * two other facts, not one this method enforces, so this branch stays in as defence in depth
-     * rather than being removed as dead code: if it ever weakens, returning whichever candidate a
-     * {@code ConcurrentHashMap} iteration reaches first would let a genuine reply complete the
-     * <em>wrong</em> ticket — silently handing the lead something that reads like a correct answer to
-     * a delegation the worker never touched, which is worse than a failure because the lead acts on
-     * it. When more than one candidate exists, guessing is not safe: fall back to the inbox exactly
-     * as the zero-candidate case does, and let {@link #abandon} apply the eventual recovery
-     * deterministically instead.
+     * <p><strong>Ambiguous match also falls to the inbox.</strong> {@link #askAnsweredAsyncTasks} can
+     * return more than one entry — a reachable state, not a hypothetical one (see its own javadoc:
+     * an {@code fleet_ask} that lapsed with no answer, fleetd #307, frees the target for a completely fresh
+     * delegation, which can itself go on to ask-and-lapse before the first worker's real reply
+     * arrives). Returning whichever candidate a {@code ConcurrentHashMap} iteration reaches first
+     * would let a genuine reply complete the <em>wrong</em> ticket — silently handing the lead
+     * something that reads like a correct answer to a delegation the worker never touched, which is
+     * worse than a failure because the lead acts on it. When more than one candidate exists, guessing
+     * is not safe: fall back to the inbox exactly as the zero-candidate case does, and let
+     * {@link #abandon} apply the eventual recovery deterministically instead.
      *
      * <p><strong>{@code content} is required (fleetd #302).</strong> Both doors that reach this
      * method must reject a missing/blank reply the same way, so the check lives here rather than in
@@ -434,15 +450,18 @@ public final class MessageService {
             count(FleetMetrics.REPLIES, "path", "rendezvous");
             return true;                       // a live send took it — unchanged fast path
         }
-        // #137: no live rendezvous waiter, but this may be the worker's real fleet_reply resuming a
-        // turn that {@link #answer} already gave up waiting on. answer()'s own bounded wait (the
-        // primary's fleet_send{turnId} call, capped well under a minute) can time out and close its
-        // waiter long before the worker — now actually resuming real work — finishes and replies. That
-        // reply used to have nowhere to land but the session inbox, leaving the async ticket's future
-        // unresolved forever: fleet_poll{ticket} stayed PENDING until fleet_stop's abandon() forced it
-        // FAILED with a misleading "session released before it replied" reason, even though the reply
-        // had, in fact, arrived. Completing the matching ticket directly here means fleet_poll{ticket}
-        // sees the real reply instead.
+        // #137/fleetd #307: no live rendezvous waiter, but this may be the worker's real fleet_reply resuming
+        // a turn that either answer() (#137) or ask() (fleetd #307) already gave up waiting on:
+        //  - answer()'s own bounded wait (the primary's fleet_send{turnId} call, capped well under a
+        //    minute) can time out and close its waiter long before the worker — now actually resuming
+        //    real work — finishes and replies.
+        //  - ask()'s own wait for the primary can time out first, with the worker resuming on its own
+        //    and finishing unanswered.
+        // Either way that reply used to have nowhere to land but the session inbox, leaving the async
+        // ticket's future unresolved forever: fleet_poll{ticket} stayed PENDING until fleet_stop's
+        // abandon() forced it FAILED with a misleading "session released before it replied" reason,
+        // even though the reply had, in fact, arrived. Completing the matching ticket directly here
+        // means fleet_poll{ticket} sees the real reply instead.
         List<Task> candidates = askAnsweredAsyncTasks(session);
         if (candidates.size() == 1) {
             Task orphan = candidates.get(0);
@@ -473,34 +492,42 @@ public final class MessageService {
     }
 
     /**
-     * Every still-open async task on {@code target} whose {@code fleet_ask} was already answered —
-     * its {@link Task#turnId} is stamped but its {@link Task#question} was cleared by {@link #answer}
-     * — yet whose future is not resolved yet (#137). Empty if no such task exists, including the
-     * common case where {@code target}'s worker never used {@code fleet_ask} at all (a task that was
-     * never asked has {@code turnId == null}, so it can never match here and only ever completes
-     * through the ordinary rendezvous fast path in {@link #reply}).
+     * Every still-open async task on {@code target} whose worker is genuinely expected to send a
+     * real {@code fleet_reply} next with nothing left registered to catch it: either its
+     * {@code fleet_ask} was already answered — {@link Task#turnId} is stamped but {@link
+     * Task#question} was cleared by {@link #answer} — or its {@code fleet_ask} lapsed unanswered and
+     * {@link Task#askTimedOut} marks that (fleetd #307; {@link Task#turnId} is {@code null} by then, forgotten
+     * so the target is not left BUSY — see {@link Task#askTimedOut}'s own javadoc). Either way the
+     * task's future is not resolved yet. Empty if no such task exists, including the common case
+     * where {@code target}'s worker never used {@code fleet_ask} at all (a task that was never asked
+     * has both {@code turnId == null} and {@code askTimedOut == false}, so it can never match here and
+     * only ever completes through the ordinary rendezvous fast path in {@link #reply}).
      *
-     * <p><strong>Returns at most one entry today — verified, not assumed.</strong> {@link #send}
-     * refuses to open a waiter on {@code target} while {@link #hasAsyncQuestion} is true, and that
-     * check matches ANY task whose {@code turnId} is still stamped in {@code asyncTasksByTurn} —
-     * not only while its question is still open. {@link #answer} deliberately leaves that stamp in
-     * place ({@code clearAsyncQuestion(turnId, false)}) until the resumed turn's own future actually
-     * resolves, at which point {@link #finishAsyncTask} both removes the stamp AND completes that
-     * task's future in the same call. So a second task can never reach "{@code turnId} stamped, future
-     * still open" — the exact pair this method matches on — while a first one already holds it: by
-     * the time the stamp is gone, so is the eligibility. This is an emergent property of those two
-     * facts holding together, not something this method (or its callers) enforces on its own — flip
-     * {@code forgetTurn} to {@code true} in that one {@link #answer} call and it silently stops being
-     * true, with nothing left to fail loudly. The callers below still handle "more than one" as
-     * defence in depth against exactly that, not because they exercise it today: {@link #reply}
-     * treats it as unresolvable and falls back to the inbox; {@link #abandon} would pick the oldest
-     * deterministically (its own {@code matching} list has no such guarantee — see its javadoc).
+     * <p><strong>Can return more than one entry — reachable, not just defence in depth.</strong>
+     * {@link #send} refuses to open a waiter on {@code target} while {@link #hasAsyncQuestion} is
+     * true, and that check matches ANY task whose {@code turnId} is still stamped in
+     * {@code asyncTasksByTurn}. While a task's {@code turnId} stays stamped — {@link #answer} leaves
+     * it in place ({@code clearAsyncQuestion(turnId, false)}) until {@link #finishAsyncTask} removes
+     * the stamp and completes the future in the same call — no second task on the same target can
+     * reach an eligible state, because {@link #send} would refuse it as BUSY first. That single-task
+     * guarantee holds only for the {@code turnId}-stamped half of this method's match: an
+     * {@link Task#askTimedOut} task is, by construction, no longer stamped in {@code asyncTasksByTurn}
+     * (that is the whole point of forgetting {@code turnId} in {@link #clearAsyncQuestion}), so the
+     * target is free the moment one ask lapses. A fresh, independent {@code sendAsync} to the same
+     * target can then be dispatched, itself pause on {@code fleet_ask}, and itself time out — landing
+     * a second {@code askTimedOut} task on the very target the first one is still waiting to answer
+     * for. Two (or more) genuinely open tasks on one target is therefore a real, reachable state
+     * today, not a hypothetical: {@link #reply} treats it as unresolvable and falls back to the
+     * inbox rather than guess which task a reply belongs to (guessing wrong would hand the lead a
+     * plausible-looking answer to a delegation the worker never touched — worse than a failure,
+     * because the lead acts on it); {@link #abandon} instead picks the oldest deterministically (its
+     * own {@code matching} list has a different, wider match — see its javadoc).
      */
     private List<Task> askAnsweredAsyncTasks(String target) {
         List<Task> candidates = new ArrayList<>();
         for (Task task : tasks.values()) {
-            if (target.equals(task.target) && task.question == null && task.turnId != null
-                    && !task.future.isDone()) {
+            if (target.equals(task.target) && task.question == null && !task.future.isDone()
+                    && (task.turnId != null || task.askTimedOut)) {
                 candidates.add(task);
             }
         }
@@ -898,6 +925,12 @@ public final class MessageService {
             return new AskResult(AskOutcome.ANSWERED, answer);
         } catch (TimeoutException e) {
             log.debug("fleet_ask from {} went unanswered in {}ms", workerSession, timeoutMillis);
+            // fleetd #307: mark the task BEFORE clearAsyncQuestion(forgetTurn=true) below drops it out of
+            // asyncTasksByTurn and nulls its turnId — that forgetting is deliberate and stays (it is
+            // what keeps the target from staying BUSY forever), but it would otherwise also erase
+            // askAnsweredAsyncTasks' only signal that the worker's eventual real fleet_reply still
+            // belongs to this task, stranding it in the inbox with a false "never replied" verdict.
+            markAskTimedOut(ticket.turnId());
             clearAsyncQuestion(ticket.turnId(), true);
             return new AskResult(AskOutcome.TIMED_OUT, null);
         } catch (ExecutionException e) {
@@ -1162,6 +1195,23 @@ public final class MessageService {
         return task;
     }
 
+    /**
+     * Mark {@code turnId}'s task as having a {@code fleet_ask} that lapsed with no answer (fleetd #307), so
+     * {@link #askAnsweredAsyncTasks} still recognizes the worker's eventual real {@code fleet_reply}
+     * as belonging to it after {@link #clearAsyncQuestion}'s {@code forgetTurn=true} erases
+     * {@link Task#turnId} — see {@link Task#askTimedOut}. Must be called before that forgetting, while
+     * {@code turnId} can still resolve the task in {@code asyncTasksByTurn}; a lookup afterward would
+     * find nothing. Only when it matches the task's current turn — same guard as
+     * {@link #clearAsyncQuestion} — so a chained second {@code fleet_ask} (#282) that already moved
+     * the task to a fresh {@code turnId} cannot mark it for a turn that is no longer its own.
+     */
+    private void markAskTimedOut(String turnId) {
+        Task task = asyncTasksByTurn.get(turnId);
+        if (task != null && turnId.equals(task.turnId)) {
+            task.askTimedOut = true;
+        }
+    }
+
     /** Clear an answered or lapsed question, but only when it matches the ticket's current turn. */
     private void clearAsyncQuestion(String turnId, boolean forgetTurn) {
         // CB-582: tell the push loop first — like ticketCollected, a removal for a turnId it never

fleetd/src/main/java/dev/ltms/fleet/placement/FixedPlacementPolicy.java

@@ -5,10 +5,14 @@ import java.util.List;
 
 /**
  * Backward-compatible placement: an unqualified spawn always resolves to the configured default
- * profile, exactly as {@code CompositePeerLauncher} did before CB-518. This ignores caps and
- * reachability so that a pre-existing config behaves identically after upgrade.
+ * profile, exactly as {@code CompositePeerLauncher} did before CB-518. This ignores caps
+ * ({@code maxLoad}) so that a pre-existing config behaves identically after upgrade — capacity
+ * gating for automatic placement is deliberately out of scope for {@code fixed}, exactly as it
+ * always has been. Reachability is a narrower exception (fleetd #315, below): a profile is never
+ * checked for reachability up front, only skipped once it has already failed in <em>this same</em>
+ * spawn call's retry loop — see the unreachable case below.
  *
- * <p>Three exceptions walk past the default instead of returning it unconditionally:
+ * <p>Four exceptions walk past the default instead of returning it unconditionally:
  * <ul>
  *   <li>Quarantine (CB-578 stage B): a quarantined default is a credential that just refused on
  *       a usage limit, not a transient capacity or reachability concern.
@@ -16,13 +20,21 @@ import java.util.List;
  *       ({@code BackendOutagePolicy}) — a separate, shorter-lived source from quarantine. When a
  *       profile is both quarantined and cooling off, only the quarantine reason is reported
  *       (exhaustion takes priority), matching {@code CompositePeerLauncher}'s explicit-spawn order.
+ *   <li>Unreachable (fleetd #315): {@code CompositePeerLauncher.spawn} retries a failed candidate
+ *       on the next one and rebuilds the {@link PlacementContext} so {@code ctx.unreachable()}
+ *       names every profile that already failed with {@code PeerUnreachableException} in this same
+ *       call. Without this check {@code select} kept handing back the same dead default forever —
+ *       the retry loop's own comment says "so the policy excludes this profile", and this is what
+ *       makes that true for {@code fixed} too, matching {@code weighted}/{@code round-robin}
+ *       (both filter on {@code ctx.unreachable()} via {@link PlacementPolicyUtil#available}).
  *   <li>Weight 0 (CB-554): {@code fixed} is still automatic selection, so a profile the operator
  *       marked "never auto-select me" ({@code weight <= 0}) must be skipped here exactly as
  *       {@code weighted}/{@code round-robin} skip it — an explicit {@code fleet_spawn} naming
  *       the profile is unaffected, only this automatic fallback walk.
  * </ul>
- * A fleet where nothing is ever quarantined, cooling off, or weight-0 never exercises any of these
- * paths, so today's behaviour is unchanged.
+ * A fleet where nothing is ever quarantined, cooling off, unreachable, or weight-0 never exercises
+ * any of these paths, so today's behaviour is unchanged — in particular, the very first selection
+ * of a spawn call always sees an empty {@code unreachable} set, so the first choice is untouched.
  */
 final class FixedPlacementPolicy implements PlacementPolicy {
 
@@ -30,12 +42,12 @@ final class FixedPlacementPolicy implements PlacementPolicy {
     public PlacementCandidate select(PlacementContext ctx) {
         String d = ctx.defaultProfile();
         if (d != null && !d.isBlank() && !ctx.quarantined().contains(d) && !ctx.coolingOff().contains(d)
-                && !weightExcluded(ctx, d)) {
+                && !ctx.unreachable().contains(d) && !weightExcluded(ctx, d)) {
             return new PlacementCandidate(d, null, 1.0f, null);
         }
         for (PlacementCandidate c : ctx.candidates()) {
             if (!ctx.quarantined().contains(c.profile()) && !ctx.coolingOff().contains(c.profile())
-                    && !c.excluded()) {
+                    && !ctx.unreachable().contains(c.profile()) && !c.excluded()) {
                 return new PlacementCandidate(c.profile(), null, c.weight(), c.maxLoad());
             }
         }
@@ -44,8 +56,9 @@ final class FixedPlacementPolicy implements PlacementPolicy {
             // Exhaustion quarantine takes priority: reported only when quarantine is absent, so the
             // message never claims "cooling off" for a profile that is really backend-exhausted.
             boolean dCoolingOff = !dQuarantined && ctx.coolingOff().contains(d);
+            boolean dUnreachable = ctx.unreachable().contains(d);
             boolean dWeightExcluded = weightExcluded(ctx, d);
-            if (dQuarantined || dCoolingOff || dWeightExcluded) {
+            if (dQuarantined || dCoolingOff || dUnreachable || dWeightExcluded) {
                 List<String> reasons = new ArrayList<>();
                 if (dQuarantined) {
                     reasons.add("is quarantined (backend exhausted)");
@@ -53,6 +66,9 @@ final class FixedPlacementPolicy implements PlacementPolicy {
                 if (dCoolingOff) {
                     reasons.add("is cooling off after repeated backend errors");
                 }
+                if (dUnreachable) {
+                    reasons.add("is unreachable");
+                }
                 if (dWeightExcluded) {
                     reasons.add("has weight 0 (excluded from automatic selection)");
                 }
@@ -62,7 +78,7 @@ final class FixedPlacementPolicy implements PlacementPolicy {
         }
         if (!ctx.candidates().isEmpty()) {
             throw new PlacementException("all worker profiles are excluded from automatic "
-                    + "selection (quarantined, cooling off, or weight-0)");
+                    + "selection (quarantined, cooling off, unreachable, or weight-0)");
         }
         throw new PlacementException("no worker profiles configured");
     }

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

@@ -93,6 +93,9 @@ public final class GitWorktrees implements Worktrees {
     /** OS group name for {@link #shareWithGroup} (fleetd #185 stage 3); {@code null} ⇒ feature off. */
     private final String group;
     private final Consumer<String> afterWorktreeAdded;
+    /** How the initial {@code git worktree add} command runs. Package-private test seam for an
+     *  interrupted command after Git has made worktree state. */
+    private final Function<String[], String> worktreeAddRunner;
     /** How {@link #shareWithGroup}'s processes (git config / chgrp / chmod / find) actually run.
      *  Defaults to the real {@link #exec(String...)}. Package-private test seam so a unit test can
      *  prove "no group configured ⇒ zero processes spawned" and inspect exactly what a configured
@@ -132,7 +135,13 @@ public final class GitWorktrees implements Worktrees {
 
     /** Test seam combining a configurable {@code group} with {@link #afterWorktreeAdded}. */
     GitWorktrees(String configuredRoot, String group, Consumer<String> afterWorktreeAdded) {
-        this(configuredRoot, group, afterWorktreeAdded, null);
+        this(configuredRoot, group, afterWorktreeAdded, null, null);
+    }
+
+    /** Test seam for changing how {@link #shareWithGroup}'s processes run. */
+    GitWorktrees(String configuredRoot, String group, Consumer<String> afterWorktreeAdded,
+                  Function<String[], String> shareGroupRunner) {
+        this(configuredRoot, group, afterWorktreeAdded, shareGroupRunner, null);
     }
 
     /**
@@ -141,13 +150,15 @@ public final class GitWorktrees implements Worktrees {
      * exactly what commands a configured group runs, without a real second OS user/group.
      *
      * @param shareGroupRunner {@code null} ⇒ the real {@link #exec(String...)}.
+     * @param worktreeAddRunner {@code null} ⇒ the real {@link #exec(String...)}.
      */
     GitWorktrees(String configuredRoot, String group, Consumer<String> afterWorktreeAdded,
-                 Function<String[], String> shareGroupRunner) {
+                  Function<String[], String> shareGroupRunner, Function<String[], String> worktreeAddRunner) {
         this.configuredRoot = configuredRoot;
         this.group = (group == null || group.isBlank()) ? null : group;
         this.afterWorktreeAdded = afterWorktreeAdded == null ? _ -> {} : afterWorktreeAdded;
         this.shareGroupRunner = shareGroupRunner != null ? shareGroupRunner : this::exec;
+        this.worktreeAddRunner = worktreeAddRunner != null ? worktreeAddRunner : this::exec;
     }
 
     @Override
@@ -166,8 +177,8 @@ public final class GitWorktrees implements Worktrees {
         String wt = path.toAbsolutePath().toString();
         log.info("adding worktree branch={} path={} base={}", branch, wt, base);
         removeUserInfoFromHttpsOrigin(repoRoot);
-        exec("git", "-C", repoRoot, "worktree", "add", wt, "-b", branch, base);
         try {
+            worktreeAddRunner.apply(new String[] {"git", "-C", repoRoot, "worktree", "add", wt, "-b", branch, base});
             afterWorktreeAdded.accept(wt);
             requireCredentialFreeHttpsOrigin(wt);
             configureEnvironmentCredentialHelper(repoRoot, wt);
@@ -181,10 +192,11 @@ public final class GitWorktrees implements Worktrees {
     }
 
     /**
-     * {@code add()} has already created the worktree and its branch by the time any step from
-     * {@link #afterWorktreeAdded} through {@link #isolateToolSurface} can throw — including
+     * {@code git worktree add} may have created the worktree and its branch by the time it, or any
+     * later step through {@link #isolateToolSurface}, throws. This includes
      * {@link #requireCredentialFreeHttpsOrigin}, an intended security refusal, not only an IO
-     * accident. Without this, {@code add()} never returns, so its caller
+     * accident. A Git-reported {@code worktree add} failure usually creates nothing, but an
+     * interrupted command can leave partial state. Without cleanup, {@code add()} never returns, so its caller
      * ({@code SessionManager#acquireWithWorktree}) never receives a path to register or clean up:
      * its local {@code path} stays null, the {@code if (path != null)} guard in its own catch block
      * never runs, and the worktree directory and branch leak on disk forever with nothing tracking
@@ -204,6 +216,10 @@ public final class GitWorktrees implements Worktrees {
      * used in {@code SessionManager#acquireWithWorktree}'s own catch block.
      */
     private void cleanupAfterAddFailure(String repoRoot, String worktreePath, String branch, RuntimeException original) {
+        if (!Files.exists(Path.of(worktreePath))) {
+            log.debug("provisioning failed before worktree {} existed; nothing to clean up", worktreePath);
+            return;
+        }
         log.warn("provisioning failed for branch={} path={}: {} — cleaning up before rethrowing",
                 branch, worktreePath, original.getMessage());
         try {

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

@@ -62,6 +62,8 @@ public final class SessionManager implements TurnListener {
     private final LongSupplier nowNanos;
     private final int contextCap;
     private final boolean clearAfterTurn;
+    /** Null in production; test seam for the interval before an idle session's conditional release. */
+    private final Consumer<MemberSession> beforeIdleRelease;
     private volatile MemberLifecycle memberLifecycle = MemberLifecycle.NONE;
     /**
      * CB-586: the repo root the fleet actually works in, remembered the first time a worktree
@@ -113,13 +115,23 @@ public final class SessionManager implements TurnListener {
     }
 
     public SessionManager(PeerLauncher launcher, Worktrees worktrees, LongSupplier nowNanos,
-                          int contextCap, boolean clearAfterTurn) {
+                           int contextCap, boolean clearAfterTurn) {
+        this(launcher, worktrees, nowNanos, contextCap, clearAfterTurn, null);
+    }
+
+    /**
+     * Package-private constructor for a deterministic reap/delivery race test. Production callers
+     * use the constructor above, whose null hook adds no callback or lock to an ordinary reap.
+     */
+    SessionManager(PeerLauncher launcher, Worktrees worktrees, LongSupplier nowNanos,
+                   int contextCap, boolean clearAfterTurn, Consumer<MemberSession> beforeIdleRelease) {
         this.launcher = launcher;
         this.worktrees = worktrees;
         this.presence = new PresenceFleet(this);
         this.nowNanos = nowNanos;
         this.contextCap = contextCap;
         this.clearAfterTurn = clearAfterTurn;
+        this.beforeIdleRelease = beforeIdleRelease;
     }
 
     /**
@@ -291,10 +303,32 @@ public final class SessionManager implements TurnListener {
      */
     private void release(String paneId, ReleaseCause cause) {
         MemberSession removed = registry.remove(paneId);
+        releaseRemoved(paneId, removed, handles.remove(paneId), cause);
+    }
+
+    /**
+     * Tear a session down only while {@code expected} is still its registry value. A lifecycle
+     * transition replaces the immutable record, so this prevents a reap based on an old READY or
+     * DONE record from stopping a worker that delivery has made BUSY.
+     */
+    private boolean releaseIfCurrent(MemberSession expected, ReleaseCause cause) {
+        if (!registry.remove(expected.paneId(), expected)) {
+            // A lifecycle transition replaced the record between the caller's check and this remove.
+            // Log it: this race is by definition unobservable otherwise, and a reaper that silently
+            // declines to reap is the hardest kind of behaviour to diagnose after the fact.
+            log.debug("skipping reap of pane={}: its registry record changed after the idle check "
+                    + "(most likely a delivery made it BUSY)", expected.paneId());
+            return false;
+        }
+        releaseRemoved(expected.paneId(), expected, handles.remove(expected.paneId()), cause);
+        return true;
+    }
+
+    private void releaseRemoved(String paneId, MemberSession removed, PeerHandle removedHandle,
+                                ReleaseCause cause) {
         // fleetd #209: remove right alongside the registry entry so a released session's handle is
         // never leaked — but keep the local reference below, so the id can still be resolved for
         // the ReleaseDetail this teardown notifies with.
-        PeerHandle removedHandle = handles.remove(paneId);
         boolean preserveWorktree = cause == ReleaseCause.SHUTDOWN;
         String snapshotRef = null;
         if (removed != null) {
@@ -865,14 +899,18 @@ public final class SessionManager implements TurnListener {
             }
             long idleNanos = now - s.lastActivityAtNanos();
             if (idleNanos > idleTtlNanos) {
-                log.debug("reaping idle session terminal={} pane={}: idle {}s exceeds the {}s ttl",
-                        s.terminalId(), s.paneId(), TimeUnit.NANOSECONDS.toSeconds(idleNanos),
-                        TimeUnit.NANOSECONDS.toSeconds(idleTtlNanos));
                 // CB-581: one session that fails to release must not abort the whole reaping pass —
                 // match drainAll's per-session try/catch so the rest of the roster still gets reaped.
                 try {
-                    release(s.paneId());
-                    reaped++;
+                    if (beforeIdleRelease != null) {
+                        beforeIdleRelease.accept(s);
+                    }
+                    if (releaseIfCurrent(s, ReleaseCause.COMPLETED)) {
+                        log.debug("reaping idle session terminal={} pane={}: idle {}s exceeds the {}s ttl",
+                                s.terminalId(), s.paneId(), TimeUnit.NANOSECONDS.toSeconds(idleNanos),
+                                TimeUnit.NANOSECONDS.toSeconds(idleTtlNanos));
+                        reaped++;
+                    }
                 } catch (RuntimeException e) {
                     log.warn("reap failed for pane={} terminal={} worktree={}; continuing with "
                             + "remaining sessions", s.paneId(), s.terminalId(), s.worktree(), e);

fleetd/src/test/java/dev/ltms/fleet/auth/CallerResolverTest.java

@@ -103,6 +103,54 @@ class CallerResolverTest {
         assertEquals(Role.PRIMARY, p.role(), "the historical behaviour, now an explicit choice");
     }
 
+    // ── fleetd #317: an unresolvable caller must never be promoted to the primary ──────────────────
+    // #305 closed the trigger where a resolved pid matched no pane *and* had no ancestry walk to
+    // save it. This is the other trigger PaneLocator's javadoc names: the pid never resolves at
+    // all — LsofPeerPidLookup returns -1 on any failure, including (silently) "lsof found no
+    // match" — so there is no candidate pid for an ancestry walk to even attempt.
+
+    /**
+     * The failing-without-the-fix case. Before #317's fix, {@code c.terminal() == null} was the
+     * only test in the loopback-trust fallback, and an unresolved pid produces exactly that same
+     * {@code null} terminal as a genuine primary — so this caller was handed
+     * {@code Principal.primary(...)}, a real worker's failed lookup becoming indistinguishable from
+     * the lead.
+     */
+    @Test
+    void aFailedPeerPidLookupIsRefusedNotPromotedToPrimary() {
+        ConnectionIdentity unresolved = new ConnectionIdentity(new PaneLocator(herdr), _ -> -1);
+        Principal p = new CallerResolver(unresolved).resolve("127.0.0.1", 55555, null);
+
+        assertEquals(Role.ANONYMOUS, p.role(),
+                "an unresolvable caller must never be silently promoted to the primary");
+    }
+
+    /**
+     * The companion invariant #317 must not break: a caller whose lookup genuinely succeeded, and
+     * who simply owns no herdr pane — the real primary's own connection — is still the primary.
+     * This is {@link #loopbackTrustTreatsANonWorkerLoopbackCallerAsThePrimary} pinned again here,
+     * named for #317 and placed next to the test it must be distinguished from: same {@code null}
+     * terminal, opposite verdict, because {@code Caller.resolved()} tells them apart.
+     */
+    @Test
+    void aRealPidThatOwnsNoPaneIsStillThePrimaryNotRefused() {
+        Principal p = new CallerResolver(nonWorkerIdentity()).resolve("127.0.0.1", 55555, null);
+
+        assertEquals(Role.PRIMARY, p.role());
+    }
+
+    /** #317 point 4: token mode never consults {@code c.pid()}, so a failed lookup must not change it. */
+    @Test
+    void tokenModeIsUndisturbedByAnUnresolvedLookup() {
+        ConnectionIdentity unresolved = new ConnectionIdentity(new PaneLocator(herdr), _ -> -1);
+        CallerResolver r = new CallerResolver(unresolved, true, "s3cret");
+
+        assertEquals(Role.ANONYMOUS, r.resolve("127.0.0.1", 55555, null).role(),
+                "no credential is still just ANONYMOUS, as before #317 — unchanged by the lookup failing");
+        assertEquals(Role.PRIMARY, r.resolve("127.0.0.1", 55555, "Bearer s3cret").role(),
+                "a valid token still authenticates the primary even though the peer-pid lookup failed");
+    }
+
     @Test
     void tokenModeRefusesANonWorkerCallerThatPresentsNoToken() {
         Principal p = new CallerResolver(nonWorkerIdentity(), true, "s3cret")

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

@@ -0,0 +1,187 @@
+package dev.ltms.fleet.config;
+
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Constructor;
+import java.lang.reflect.RecordComponent;
+import java.util.Arrays;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.TreeSet;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * fleetd #323: {@code ConfigRef.sameLaunchSettings} javadoc used to claim it "compares every
+ * component the launcher reads at spawn". It did not — {@code ideProjectDir}, {@code
+ * ideOpenCommand} and {@code autoCompactWindow} were all baked in at daemon startup (see
+ * {@code ClaudeCodeLauncher}/{@code OpenCodeLauncher}) and missing from the comparison, so a reload
+ * that changed only one of them reported "config reloaded" and the running daemon kept the old
+ * value.
+ *
+ * <p>This class is the mechanism the issue asked for: it enumerates every record component of
+ * {@link FleetConfig.Profile} by reflection and proves — by actually mutating a base profile one
+ * field at a time and calling the real method — that each component is either compared by
+ * {@link ConfigRef#sameLaunchSettings} or named in {@link ConfigRef#LAUNCH_SETTINGS_EXCLUDED} with
+ * a reason. A new profile field that is neither fails this test by name, not a hand-maintained list
+ * going stale.
+ */
+class ConfigRefProfileCoverageTest {
+
+    private static final RecordComponent[] COMPONENTS = FleetConfig.Profile.class.getRecordComponents();
+
+    /**
+     * One valid, non-blank value per record component — "the a value". None of these trip any
+     * defaulting/normalization in {@code Profile}'s compact constructor (see {@code
+     * FleetConfig.java}), so what goes in is what {@code sameLaunchSettings} sees back out.
+     */
+    private static final Map<String, Object> BASE = baseValues();
+
+    /** The same shape, each value distinct from {@link #BASE} — "the b value". */
+    private static final Map<String, Object> ALT = altValues();
+
+    private static Map<String, Object> baseValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        v.put("profile", "sonnet");
+        v.put("baseUrl", "http://gx00.gw:8000");
+        v.put("model", "sonnet");
+        v.put("configDir", "/config/a");
+        v.put("tokenEnv", "TOKEN_A");
+        v.put("argv", List.of("claude", "--flag-a"));
+        v.put("placement", "tab");
+        v.put("workspace", "workspace-a");
+        v.put("tabLabel", "label-a");
+        v.put("mcpUrl", "http://mcp-a");
+        v.put("cwd", "/cwd/a");
+        v.put("parityOverlay", List.of(".env", ".env.a"));
+        v.put("gitTokenEnv", "GIT_TOKEN_A");
+        v.put("gitHostEnv", "GITEA_HOST_A");
+        v.put("kind", "claude-code");
+        v.put("env", Map.of("K", "A"));
+        v.put("weight", 1.0f);
+        v.put("maxLoad", 5);
+        v.put("subscription", Boolean.TRUE);
+        v.put("exhaustedPattern", "usage limit a");
+        v.put("credentialId", "cred-a");
+        v.put("ideMcpUrl", "http://ide-mcp-a");
+        v.put("ideProjectDir", "modules/a");
+        v.put("ideOpenCommand", "open-cmd-a {dir}");
+        v.put("autoCompactWindow", 150000);
+        v.put("errorPattern", "error a");
+        assertNamesMatchComponents(v);
+        return v;
+    }
+
+    private static Map<String, Object> altValues() {
+        Map<String, Object> v = new LinkedHashMap<>();
+        v.put("profile", "sonnet-b");
+        v.put("baseUrl", "http://gx01.gw:8000");
+        v.put("model", "haiku");
+        v.put("configDir", "/config/b");
+        v.put("tokenEnv", "TOKEN_B");
+        v.put("argv", List.of("claude", "--flag-b"));
+        v.put("placement", "weighted");
+        v.put("workspace", "workspace-b");
+        v.put("tabLabel", "label-b");
+        v.put("mcpUrl", "http://mcp-b");
+        v.put("cwd", "/cwd/b");
+        v.put("parityOverlay", List.of(".env", ".env.b"));
+        v.put("gitTokenEnv", "GIT_TOKEN_B");
+        v.put("gitHostEnv", "GITEA_HOST_B");
+        v.put("kind", "opencode");
+        v.put("env", Map.of("K", "B"));
+        v.put("weight", 2.0f);
+        v.put("maxLoad", 9);
+        v.put("subscription", Boolean.FALSE);
+        v.put("exhaustedPattern", "usage limit b");
+        v.put("credentialId", "cred-b");
+        v.put("ideMcpUrl", "http://ide-mcp-b");
+        v.put("ideProjectDir", "modules/b");
+        v.put("ideOpenCommand", "open-cmd-b {dir}");
+        v.put("autoCompactWindow", 250000);
+        v.put("errorPattern", "error b");
+        assertNamesMatchComponents(v);
+        return v;
+    }
+
+    private static void assertNamesMatchComponents(Map<String, Object> values) {
+        Set<String> componentNames = new TreeSet<>();
+        for (RecordComponent rc : COMPONENTS) {
+            componentNames.add(rc.getName());
+        }
+        assertEquals(componentNames, new TreeSet<>(values.keySet()),
+                "this test's value map has drifted from FleetConfig.Profile's actual components — "
+                        + "update BASE/ALT alongside the record");
+    }
+
+    private static FleetConfig.Profile profileOf(Map<String, Object> values) throws ReflectiveOperationException {
+        Class<?>[] types = Arrays.stream(COMPONENTS).map(RecordComponent::getType).toArray(Class<?>[]::new);
+        Object[] args = Arrays.stream(COMPONENTS)
+                .map(rc -> values.get(rc.getName()))
+                .toArray();
+        Constructor<FleetConfig.Profile> ctor = FleetConfig.Profile.class.getDeclaredConstructor(types);
+        return ctor.newInstance(args);
+    }
+
+    /** {@code BASE} with exactly one named component swapped for its {@code ALT} value. */
+    private static FleetConfig.Profile mutate(String componentName) throws ReflectiveOperationException {
+        Map<String, Object> values = new LinkedHashMap<>(BASE);
+        values.put(componentName, ALT.get(componentName));
+        return profileOf(values);
+    }
+
+    /**
+     * The mechanism fleetd #323 asked for: enumerate {@link FleetConfig.Profile}'s record
+     * components, mutate each non-excluded one, and prove {@code sameLaunchSettings} actually
+     * notices — not just that some hand-maintained list claims it does. Prints the denominator
+     * (total / compared / excluded) the issue required: a checker that cannot state its own
+     * denominator is the failure this repo keeps hitting.
+     */
+    @Test
+    void sameLaunchSettingsComparesEveryProfileComponentOrExcludesIt() throws ReflectiveOperationException {
+        int total = COMPONENTS.length;
+        Set<String> excluded = ConfigRef.LAUNCH_SETTINGS_EXCLUDED;
+
+        Set<String> allNames = new TreeSet<>();
+        for (RecordComponent rc : COMPONENTS) {
+            allNames.add(rc.getName());
+        }
+        assertTrue(allNames.containsAll(excluded),
+                "ConfigRef.LAUNCH_SETTINGS_EXCLUDED names a component that does not exist on "
+                        + "FleetConfig.Profile — check for a typo: " + excluded);
+
+        FleetConfig.Profile base = profileOf(BASE);
+        List<String> uncovered = new java.util.ArrayList<>();
+        int compared = 0;
+        for (RecordComponent rc : COMPONENTS) {
+            String name = rc.getName();
+            if (excluded.contains(name)) {
+                continue;
+            }
+            FleetConfig.Profile mutated = mutate(name);
+            if (ConfigRef.sameLaunchSettings(base, mutated)) {
+                uncovered.add(name);
+            } else {
+                compared++;
+            }
+        }
+
+        System.out.printf(
+                "ConfigRef.sameLaunchSettings coverage — %d Profile components total, %d compared, "
+                        + "%d excluded (%s)%n",
+                total, compared, excluded.size(), excluded);
+
+        assertEquals(List.of(), uncovered,
+                "these FleetConfig.Profile components changed but ConfigRef.sameLaunchSettings "
+                        + "reported no difference — add each one to the comparison (it is read at "
+                        + "spawn and baked in until a restart) or to ConfigRef.LAUNCH_SETTINGS_EXCLUDED "
+                        + "with a reason it is genuinely read live: " + uncovered);
+        assertEquals(total, compared + excluded.size(),
+                "every FleetConfig.Profile record component must be either compared or excluded — "
+                        + total + " components, " + compared + " compared, " + excluded.size()
+                        + " excluded");
+    }
+}

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

@@ -434,6 +434,78 @@ class ConfigRefTest {
         assertEquals("provider 5xx", ref.get().profiles().get("sonnet").errorPattern());
     }
 
+    /**
+     * fleetd #323 instance 1: {@code ideProjectDir} is read at spawn off the frozen profile map
+     * (see {@code ClaudeCodeLauncher}/{@code OpenCodeLauncher}) exactly like {@code model}, but was
+     * missing from {@code sameLaunchSettings} — a reload changing only this field used to report a
+     * bare "config reloaded" and the running daemon kept launching with the old value.
+     */
+    @Test
+    void changingAProfilesIdeProjectDirIsReportedAsDeferred(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                herdrSocket: ~/.config/herdr/herdr.sock
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx00.gw:8000
+                    model: sonnet
+                    ideProjectDir: fleetd
+                guard:
+                  offSubscriptionHosts:
+                    - gx00.gw
+                """);
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, """
+                bind:
+                  host: 127.0.0.1
+                  port: 8765
+                herdrSocket: ~/.config/herdr/herdr.sock
+                profiles:
+                  sonnet:
+                    baseUrl: http://gx00.gw:8000
+                    model: sonnet
+                    ideProjectDir: fleetd-renamed
+                guard:
+                  offSubscriptionHosts:
+                    - gx00.gw
+                """);
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(1, out.deferred().size(), out.deferred().toString());
+        assertTrue(out.deferred().getFirst().contains("sonnet"), out.deferred().toString());
+        assertTrue(out.deferred().getFirst().contains("launch settings"), out.deferred().toString());
+        // The snapshot still carries the new value — a restart is what makes it take effect.
+        assertEquals("fleetd-renamed", ref.get().profiles().get("sonnet").ideProjectDir());
+    }
+
+    /**
+     * fleetd #323 instance 2: {@code worktreeGroup} is baked into the same {@code GitWorktrees}
+     * as {@code worktreeRoot} (Fleetd.java:251) and never rebuilt, but only {@code worktreeRoot}
+     * was on {@code changedDeferredKeys} — a reload changing only the group reported a bare
+     * "config reloaded" and newly provisioned worktrees kept the old sharing behaviour.
+     */
+    @Test
+    void changingWorktreeGroupIsReportedAsDeferred(@TempDir Path dir) throws Exception {
+        Path f = dir.resolve("fleetd.yaml");
+        Files.writeString(f, yaml("worktreeGroup: devgroup\n"));
+        ConfigRef ref = refFor(f);
+
+        Files.writeString(f, yaml("worktreeGroup: devgroup2\n"));
+        ConfigRef.Outcome out = ref.reload();
+
+        assertTrue(out.applied());
+        assertEquals(java.util.List.of("worktreeGroup"), out.deferred());
+        assertTrue(out.summary().contains("needs a restart") || out.summary().contains("need a restart"),
+                out.summary());
+        // The snapshot still carries the new value — a restart is what makes it take effect.
+        assertEquals("devgroup2", ref.get().worktreeGroup());
+    }
+
     @Test
     void aFixedRefHasNoFileAndRefusesToReload() {
         FleetConfig cfg = new FleetConfig(null, null, null, null, null, null,

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

@@ -43,6 +43,22 @@ class ConnectionIdentityTest {
         assertNull(with(_ -> 999_999).callerTerminal("127.0.0.1", 55555));
     }
 
+    @Test
+    void callerIsUnresolvedWhenThePeerPidLookupFails() {
+        // fleetd #317: LsofPeerPidLookup returns -1 on any failure — a fork error, or (silently)
+        // simply no matching lsof line. Caller.resolved() is the one place that sentinel is tested.
+        ConnectionIdentity.Caller c = with(_ -> -1).resolve("127.0.0.1", 55555);
+        assertFalse(c.resolved(), "a -1 pid means the lookup failed, not that this pid owns no pane");
+    }
+
+    @Test
+    void callerIsResolvedWhenThePidIsRealEvenThoughItOwnsNoPane() {
+        // The primary's own connection: a real, lsof-found pid that just isn't a worker pane. This
+        // must read as "resolved" — the distinction #317 turns on.
+        ConnectionIdentity.Caller c = with(_ -> 999_999).resolve("127.0.0.1", 55555);
+        assertTrue(c.resolved());
+    }
+
     @Test
     void resolvesTheCallersPidAndCwd() {
         // CB-112: the primary maps to no pane, but its PID and cwd are still readable.

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

@@ -189,7 +189,7 @@ class FleetMcpTest {
     }
 
     @Test
-    void unansweredAsyncAskReturnsTheTicketToPending() throws Exception {
+    void unansweredAsyncAskReturnsTheTicketToPendingThenAWorkersLateReplyStillCompletesIt() throws Exception {
         McpSchema.CallToolResult accepted = FleetMcp.sendAsync(messages, "term_a", "do it", null, Set.of());
         String ticket = textOf(accepted).substring(textOf(accepted).indexOf("ticket=") + "ticket=".length()).trim();
 
@@ -203,8 +203,15 @@ class FleetMcpTest {
         assertTrue(textOf(ask).contains("no answer"), textOf(ask));
         assertTrue(textOf(FleetMcp.poll(messages, ticket, null)).startsWith("[pending"));
 
+        // fleetd #307: the worker resumed on its own after the primary never answered, and its real
+        // fleet_reply must complete its OWN async ticket — not strand in the inbox with
+        // 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");
-        assertEquals("finished after timeout", messages.drainReplies("term_a").getFirst().content());
+        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");
     }
 
     @Test

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

@@ -615,6 +615,67 @@ class CompositePeerLauncherTest {
         assertEquals(1, adapter.spawnCount("b"));
     }
 
+    /**
+     * fleetd #315: {@code CompositePeerLauncher.spawn} rebuilds the {@link PlacementContext} after
+     * every failed attempt "so the policy excludes this profile" (see the comment at the retry call
+     * site) — but {@code FixedPlacementPolicy} never read {@code ctx.unreachable()}, so under the
+     * default {@code fixed} placement every retry re-picked the same dead default and a second,
+     * healthy, configured profile was never tried. This is the same scenario as
+     * {@link #failoverRetriesNextCandidateWhenProfileIsUnreachable}, but pinned to {@code fixed()}
+     * instead of {@code weighted()} — the three existing failover tests all use {@code weighted()},
+     * which is exactly why nobody caught this: the retry loop's contract has no coverage under its
+     * own default policy.
+     */
+    @Test
+    void failoverRetriesNextCandidateUnderFixedPlacementWhenProfileIsUnreachable() {
+        FakeHerdr herdr = new FakeHerdr();
+        Map<String, FleetConfig.Profile> profiles = ordered(
+                "a", stubWorker("a"),
+                "b", stubWorker("b"));
+        StubLauncher adapter = new StubLauncher("claude", herdr, profiles, "a", Set.of("a"));
+        CompositePeerLauncher composite = new CompositePeerLauncher(
+                List.of(adapter), "a", profiles, PlacementPolicies.fixed(), _ -> 0);
+
+        PeerHandle h = composite.spawn(new SpawnRequest(null, null, null));
+        assertEquals("b", h.profile(),
+                "fixed placement must fail over from the unreachable default a to the healthy b");
+        assertEquals(1, adapter.spawnCount("a"), "a was tried once and failed");
+        assertEquals(1, adapter.spawnCount("b"), "b was tried once and succeeded");
+    }
+
+    /**
+     * fleetd #315: the same fix — {@code FixedPlacementPolicy} consulting {@code ctx.unreachable()}
+     * — also covers the wiring-bug branch in {@code CompositePeerLauncher.spawn}: a profile that
+     * placement is allowed to choose (it is in the configured candidate list) but that no delegate
+     * declares ({@code byProfile.get(chosen.profile()) == null}). That branch adds the profile to
+     * {@code unreachable} and {@code continue}s without ever calling a launcher, so before this fix
+     * {@code fixed} handed back the same adapterless profile on every remaining attempt too.
+     */
+    @Test
+    void failoverSkipsAConfiguredProfileNoAdapterDeclaresUnderFixedPlacement() {
+        FakeHerdr herdr = new FakeHerdr();
+        // Placement's candidate list has three profiles, in this order (LinkedHashMap preserves it,
+        // and the fixed default resolves to the first — see the `ordered` helper's own javadoc).
+        Map<String, FleetConfig.Profile> profiles = new LinkedHashMap<>();
+        profiles.put("c", stubWorker("c"));
+        profiles.put("a", stubWorker("a"));
+        profiles.put("b", stubWorker("b"));
+        // The adapter only declares a and b — c is a configured profile with no owning adapter,
+        // the "wiring bug" the comment in CompositePeerLauncher.spawn calls out.
+        Map<String, FleetConfig.Profile> adapterProfiles = new LinkedHashMap<>();
+        adapterProfiles.put("a", stubWorker("a"));
+        adapterProfiles.put("b", stubWorker("b"));
+        StubLauncher adapter = new StubLauncher("claude", herdr, adapterProfiles, "a", Set.of());
+        CompositePeerLauncher composite = new CompositePeerLauncher(
+                List.of(adapter), "a", profiles, PlacementPolicies.fixed(), _ -> 0);
+
+        PeerHandle h = composite.spawn(new SpawnRequest(null, null, null));
+        assertEquals("a", h.profile(),
+                "c has no adapter, so fixed placement must skip it and land on the next candidate, a");
+        assertEquals(0, adapter.spawnCount("c"), "c is never spawned — no adapter owns it");
+        assertEquals(1, adapter.spawnCount("a"));
+    }
+
     @Test
     void explicitSpawnAtMaxLoadThrowsPlacementExceptionNamingProfileLiveAndCap() {
         FakeHerdr herdr = new FakeHerdr();

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

@@ -957,6 +957,76 @@ class MessageServiceTest {
         assertEquals(MessageService.Phase.DONE, awaitTicketPhase(next, MessageService.Phase.DONE).phase());
     }
 
+    /**
+     * fleetd #307: a worker's {@code fleet_ask} can time out because the primary never answers —
+     * distinct from {@link #aReplyAfterAnswerTimesOutStillCompletesTheAsyncTicket}, where the
+     * primary DID answer and only its own bounded wait for the resumed turn expired.
+     * {@code ask()}'s timeout path deliberately forgets the task's {@code turnId} (so
+     * {@code hasAsyncQuestion} stops reporting the target BUSY — see
+     * {@code unansweredAsyncQuestionReturnsTheTicketToPendingAndReleasesItsTarget} above), which used
+     * to also erase the one signal {@code askAnsweredAsyncTasks} needed to recognize the worker's
+     * eventual real {@code fleet_reply}. That reply then had nowhere to land but the inbox, and
+     * {@code fleet_poll{ticket}} stayed PENDING forever — later force-failed with the false reason
+     * "session released before it replied", even though the worker had, in fact, replied.
+     */
+    @Test
+    void aReplyAfterAnAskTimeoutStillCompletesTheAsyncTicket() throws Exception {
+        String ticket = messages.sendAsync(T, "task that asks then finishes alone");
+        awaitWaiting();
+        injectDelivery();
+
+        assertEquals(MessageService.AskOutcome.TIMED_OUT,
+                messages.ask(T, "which config?", 200).outcome());
+        assertEquals(MessageService.Phase.PENDING, messages.poll(ticket).phase(),
+                "only the question wait ended; the delegated turn may still finish");
+
+        // The worker keeps working past the timeout and only now calls fleet_reply — with no live
+        // rendezvous waiter open (ask()'s timeout already closed it) and no new send() having
+        // reopened one for this target.
+        assertTrue(messages.reply(T, "PR opened: https://example/pulls/42"));
+
+        MessageService.TaskView done = awaitTicketPhase(ticket, MessageService.Phase.DONE);
+        assertEquals("PR opened: https://example/pulls/42", done.reply(),
+                "fleet_poll{ticket} must return the worker's real reply, not stay pending forever");
+        assertEquals("reply", done.replySource());
+        assertFalse(messages.hasStrandedReply(T),
+                "the reply completed its own ticket directly and never touched the inbox");
+    }
+
+    /**
+     * fleetd #307's ambiguity guard: an ask timeout frees its target ({@code hasAsyncQuestion}
+     * becomes false the instant it lapses — proven above), so a second, independent delegation can
+     * be dispatched to the same target and itself go on to ask-and-lapse before the first worker's
+     * real reply ever arrives. Two open tasks are then both eligible candidates on one target with
+     * no live waiter to disambiguate them. A reply arriving now must not guess which one it answers
+     * — guessing wrong would hand the lead a plausible-looking answer to a delegation the worker
+     * never touched, worse than a failure because the lead acts on it — so it must fall back to the
+     * inbox exactly as the zero-candidate case does.
+     */
+    @Test
+    void twoAskTimedOutTicketsOnOneTargetFallBackToTheInboxRatherThanGuess() throws Exception {
+        String ticket1 = messages.sendAsync(T, "first task that asks");
+        awaitWaiting();
+        injectDelivery();
+        assertEquals(MessageService.AskOutcome.TIMED_OUT, messages.ask(T, "Q1?", 200).outcome());
+
+        String ticket2 = messages.sendAsync(T, "second task that asks");
+        awaitWaiting();
+        injectDelivery();
+        assertEquals(MessageService.AskOutcome.TIMED_OUT, messages.ask(T, "Q2?", 200).outcome());
+
+        assertTrue(messages.reply(T, "which task does this answer?"));
+
+        assertEquals(MessageService.Phase.PENDING, messages.poll(ticket1).phase(),
+                "an ambiguous reply must not guess ticket1");
+        assertEquals(MessageService.Phase.PENDING, messages.poll(ticket2).phase(),
+                "an ambiguous reply must not guess ticket2");
+        assertTrue(messages.hasStrandedReply(T));
+        var drained = messages.drainReplies(T);
+        assertEquals(1, drained.size());
+        assertEquals("which task does this answer?", drained.get(0).content());
+    }
+
     @Test
     void asyncQuestionBelongsToTheTaskThatOwnsItsForwardWaiter() throws Exception {
         String first = messages.sendAsync(T, "first task");

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

@@ -403,6 +403,52 @@ class GitWorktreesTest {
         assertTrue(heads.isBlank(), "the branch leaked after a post-creation step threw:\n" + heads);
     }
 
+    /**
+     * fleetd #309. The add runner is a narrow seam for the case where Git has created state but
+     * the caller then kills the process. The runner first performs the real add in this throwaway
+     * repo, then throws the same kind of exception that {@link GitWorktrees#exec} uses for a timeout.
+     * This proves the failure path cleans both real Git objects without waiting for a slow checkout.
+     */
+    @Test
+    void addCleansUpWhenTheWorktreeAddRunnerFailsAfterCreatingState(@TempDir Path tmp) throws Exception {
+        Path repo = initRepo(tmp.resolve("repo"));
+        String branch = "cb-309-timeout";
+        WorktreeException timeout = new WorktreeException("command timed out: synthetic git worktree add");
+        AtomicReference<String> createdPath = new AtomicReference<>();
+        GitWorktrees worktrees = new GitWorktrees(tmp.resolve("wts").toString(), null, null, null, command -> {
+            createdPath.set(command[5]);
+            try {
+                git(repo, "worktree", "add", command[5], "-b", command[7], command[8]);
+            } catch (Exception e) {
+                throw new AssertionError("test setup could not create the worktree", e);
+            }
+            throw timeout;
+        });
+
+        WorktreeException thrown = assertThrows(WorktreeException.class,
+                () -> worktrees.add(repo.toString(), branch, "HEAD"));
+
+        assertSame(timeout, thrown, "cleanup must not replace the add failure");
+        assertNotNull(createdPath.get(), "the add runner must receive the worktree path");
+        assertFalse(Files.exists(Path.of(createdPath.get())),
+                "the worktree directory leaked after the add runner failed");
+        assertFalse(refExists(repo, "refs/heads/" + branch), "the branch leaked after the add runner failed");
+    }
+
+    /** An ordinary Git refusal must not delete the existing branch or log a cleanup warning. */
+    @Test
+    void addFailureBeforeCreatingAWorktreeIsQuiet(@TempDir Path tmp) throws Exception {
+        Path repo = initRepo(tmp.resolve("repo"));
+        String branch = "already-exists";
+        git(repo, "branch", branch);
+
+        assertThrows(WorktreeException.class, () -> new GitWorktrees(tmp.resolve("wts").toString())
+                .add(repo.toString(), branch, "HEAD"));
+
+        assertTrue(refExists(repo, "refs/heads/" + branch), "the existing branch must remain");
+        assertTrue(capturedMessages().isEmpty(), "an ordinary Git refusal logged a warning: " + capturedMessages());
+    }
+
     // ---- CB-189: broader remote-URL coverage — every remote, both fetch and push URLs, any
     // non-SSH scheme. Reporting only, additive to the origin/https strip-and-refuse tests above. ----
 

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

@@ -183,14 +183,19 @@ class SessionManagerTest {
     }
 
     private SessionManager sessionManager(FakeHerdr herdr, LongSupplier clock, int contextCap,
-                                          boolean clearAfterTurn) {
+                                           boolean clearAfterTurn) {
+        return sessionManager(herdr, clock, contextCap, clearAfterTurn, null);
+    }
+
+    private SessionManager sessionManager(FakeHerdr herdr, LongSupplier clock, int contextCap,
+                                           boolean clearAfterTurn, java.util.function.Consumer<MemberSession> hook) {
         FleetConfig.Profile cfg = new FleetConfig.Profile(
                 "ltms-local", "http://gx00.gw:8000", "coder", null, "FLEETD_WORKER_TOKEN",
                 List.of("ccs", "ltms-local"), "tab", "fleetd-workers",
                 "worker: {profile} #{n}", null, null, null);
         ClaudeCodeLauncher workers = new ClaudeCodeLauncher(new AgentControl(herdr), new WorkspaceControl(herdr),
                 new SubscriptionGuard(Set.of("gx00.gw")), Map.of(cfg.profile(), cfg), cfg.profile(), _ -> null);
-        return new SessionManager(workers, new GitWorktrees(), clock, contextCap, clearAfterTurn);
+        return new SessionManager(workers, new GitWorktrees(), clock, contextCap, clearAfterTurn, hook);
     }
 
     @Test
@@ -675,6 +680,24 @@ class SessionManagerTest {
                 "BUSY session remains");
     }
 
+    @Test
+    void reapIdleDoesNotReleaseSessionDeliveredAfterItsEligibilityCheck() {
+        long[] clock = {0};
+        FakeHerdr herdr = new FakeHerdr();
+        SessionManager[] manager = new SessionManager[1];
+        SessionManager sessions = sessionManager(herdr, () -> clock[0], 0, false,
+                session -> manager[0].onDelivered(session.terminalId(), TestTurnTokens.inert(session.terminalId())));
+        manager[0] = sessions;
+        MemberSession session = sessions.acquire("ltms-local", null, "/caller", "term_primary");
+        sessions.asPresence().markPresent(session.terminalId());
+
+        clock[0] = 11;
+        assertEquals(0, sessions.reapIdle(10), "delivery replaces the idle snapshot before release");
+        assertEquals(MemberSession.State.BUSY, sessions.get(session.paneId()).orElseThrow().state(),
+                "a just-delivered session stays registered and busy");
+        assertFalse(herdr.called("pane.close"), "the busy session pane is not stopped");
+    }
+
     @Test
     void doneSessionPastIdleTtlIsReaped() {
         long[] clock = {0};