feat(nomos): chat working-visibility, message queue, generation-aware timeline

Make background/long/desynced turns visible and queueable, fixing the four
symptoms that survived the v0.15.0 chat reliability pass.

F1 - status-driven working signal (workspace.ts taskWorking/currentWorking =
streaming OR status in {planning,executing}). Drives the chat trace, indicator,
and activity spinner so a turn with no live stream (background resume, a dropped
SSE, an idle-close mid long turn) still looks alive.

F2 - operator messages sent during an in-flight turn are now QUEUED and
auto-run when the gate frees, replacing the "still finishing a previous step...
send it again" rejection. Per-session in-memory FIFO (messagequeue.go, capped at
20) drained one-at-a-time under the turn gate; a `queued` SSE event drives a
"Queued" hint. drainQueued releases via a per-iteration deferred closure so a
runChatTurn panic can't deadlock the session's gate.

F3 - SSE keepalive (12s `:keepalive` comment) in handleChat so 20-40s
inter-iteration gaps no longer trip a proxy/browser idle close (the desync root
cause). All SSE writes serialized through one mutex.

F4 - generation-aware activity timeline (only the last propose_plan renders;
superseded ones collapse to one "Earlier plan revised" marker; step-attribution
follows only the current generation) + debounced plan refetch on lifecycle
events so a missed plan.proposed self-heals.

Verified against the last session (23da10db: 6m33s turn, operator "status"
deferred at 19:48:05). go test ./cmd/nomos/ green (new messagequeue tests);
web vitest 72/72 (new F4 generation tests); vite build clean.

VERSION: 0.16.0 -> 0.17.0

cmd/nomos/agent.go

@@ -60,6 +60,10 @@ type agent struct {
 	// gate serializes turns per session (at most one in-flight turn per
 	// sessionID). See turngate.go and plan 2026-08-03 F1.
 	gate *turnGate
+	// queue holds operator messages that arrived while a turn was already
+	// running; they are auto-run when the gate frees (plan 2026-08-03 F2).
+	// See messagequeue.go.
+	queue *messageQueue
 }
 
 func newAgent(ctx context.Context, clients *mcpClientPool, st *store, agentSlug string) (*agent, error) {
@@ -121,6 +125,7 @@ func newAgent(ctx context.Context, clients *mcpClientPool, st *store, agentSlug
 		apiToken:   os.Getenv("OIKOS_MCP_BEARER_TOKEN"),
 		httpClient: &http.Client{Timeout: 15 * time.Second},
 		gate:       newTurnGate(),
+		queue:      newMessageQueue(),
 	}, nil
 }
 

cmd/nomos/continue.go

@@ -222,7 +222,13 @@ func (a *agent) resumeSession(ctx context.Context, sessionID, note string) bool
 		slog.Info("nomos: turn already active, skipping background resume", "session", sessionID)
 		return false
 	}
-	defer a.gate.release(sessionID)
+	// Release the gate, then drain any operator message that was queued while
+	// this background turn ran (plan 2026-08-03 F2). Queued messages are run as
+	// real user turns server-side; resumeSession itself never enqueues.
+	defer func() {
+		a.gate.release(sessionID)
+		safego.Go("nomos:drain:"+sessionID, func() { a.drainQueued(context.Background(), sessionID) })
+	}()
 
 	placeholder, _ := json.Marshal(map[string]any{
 		"role": "assistant",

cmd/nomos/main.go

@@ -167,6 +167,139 @@ func sseEvent(w http.ResponseWriter, flusher http.Flusher, event agentEvent) {
 	flusher.Flush()
 }
 
+// runChatTurn is the shared core of an operator-initiated turn: insert an
+// assistant placeholder, run a.chat with incremental persistence (so whatever
+// happened before an abort is never lost), finalize the row, and derive a
+// title. It is agnostic to the transport: `sink` receives every agent event
+// for delivery (SSE for a live handleChat, a no-op for a queued turn that has
+// no client attached — the frontend learns about those via the poller + the
+// status-driven "working" signal). The caller MUST already hold the session's
+// turn-gate permit.
+func (a *agent) runChatTurn(pctx, ctx context.Context, sessionID, message string, sink func(agentEvent)) {
+	toolCalls := []map[string]any{}
+	// P3: accumulate per-iteration reasoning instead of overwriting with the
+	// final `text` event (see the original inline comment in handleChat).
+	var textParts []string
+	var finalText string
+
+	placeholder, _ := json.Marshal(map[string]any{"role": "assistant", "text": ""})
+	msgID, err := a.store.insertMessageReturningID(pctx, sessionID, "assistant", placeholder)
+	if err != nil {
+		slog.Error("nomos: chat placeholder insert failed", "session", sessionID, "error", err)
+	}
+	persist := func() {
+		if msgID == uuid.Nil {
+			return
+		}
+		body, _ := json.Marshal(map[string]any{
+			"role":       "assistant",
+			"text":       finalText,
+			"tool_calls": toolCalls,
+		})
+		a.store.updateMessage(pctx, msgID, body)
+	}
+
+	a.chat(ctx, sessionID, message, func(ev agentEvent) {
+		if ev.Type == "tool_use" || ev.Type == "tool_result" {
+			if m, ok := ev.Data.(map[string]any); ok {
+				m["type"] = ev.Type
+				// One entry per tool call: tool_use creates it, tool_result
+				// merges the result into the same entry (matched by id).
+				id, _ := m["id"].(string)
+				if id != "" && ev.Type == "tool_result" {
+					for _, existing := range toolCalls {
+						if eID, _ := existing["id"].(string); eID == id {
+							for k, v := range m {
+								existing[k] = v
+							}
+							break
+						}
+					}
+				} else {
+					toolCalls = append(toolCalls, m)
+				}
+			}
+			persist() // live: survives even if the client disconnects right after
+		}
+		if ev.Type == "text" {
+			if t, ok := ev.Data.(string); ok && t != "" {
+				textParts = append(textParts, t)
+				finalText = strings.Join(textParts, "\n\n")
+				persist()
+			}
+		}
+		sink(ev)
+	})
+
+	// B.6: if the turn ended with no text and no tool calls (the model
+	// empty-response'd and all retries failed), delete the placeholder row
+	// instead of persisting an empty bubble.
+	if finalText == "" && len(toolCalls) == 0 && msgID != uuid.Nil {
+		a.store.deleteMessage(pctx, msgID)
+	} else {
+		persist() // final state — same row, updated one last time
+	}
+
+	// Title: prefer the goal once set; else the first assistant answer.
+	if finalText != "" && sessionID != "ephemeral" {
+		var goalTitle string
+		if sess, gerr := a.store.getSession(pctx, sessionID); gerr == nil && sess.Goal != "" {
+			goalTitle = truncate(sess.Goal, 120)
+		}
+		title := goalTitle
+		if title == "" {
+			title = truncate(finalText, 80)
+		}
+		if title != "" {
+			a.store.updateSessionTitle(pctx, sessionID, title)
+		}
+	}
+}
+
+// drainAcquireWait is how long drainQueued blocks for a busy gate before
+// re-queuing and deferring to the holder's own release-drain. A package var so
+// tests can shorten it; in production it just needs to outlast the brief
+// release→drain handoff window.
+var drainAcquireWait = 5 * time.Second
+
+// drainQueued runs every queued operator message for a session as its own turn,
+// one at a time, under the turn gate. Called (in a goroutine) whenever a turn
+// releases the gate — from handleChat (live) and resumeSession (background) —
+// so a message queued while the agent was busy is acted on as soon as it's
+// free, without the operator re-sending. See messagequeue.go (plan 2026-08-03
+// F2).
+//
+// Each queued turn is persisted incrementally and has no SSE client (the
+// browser detached after receiving the `queued` event); the frontend sees the
+// result via the 3s poller and the status-driven "working" indicator.
+func (a *agent) drainQueued(ctx context.Context, sessionID string) {
+	for {
+		msg, ok := a.queue.dequeue(sessionID)
+		if !ok {
+			return
+		}
+		// Block briefly for the gate. If a live turn grabbed it first, put the
+		// message back — that turn's release will drain it again. Never stack.
+		if !a.gate.acquire(sessionID, drainAcquireWait) {
+			a.queue.requeueFront(sessionID, msg)
+			return
+		}
+		slog.Info("nomos: running queued operator message", "session", sessionID)
+		pctx := context.Background()
+		// Run the turn inside a per-iteration closure so the gate release is
+		// deferred to the end of THIS turn (and runs even if runChatTurn
+		// panics — safego recovers the panic at the goroutine boundary, so a
+		// non-deferred release would be skipped and the session's permit held
+		// forever, deadlocking all future turns). A bare `defer release` in
+		// the loop would be wrong too: Go defers run at function exit, not
+		// iteration exit, so the gate would stay held across iterations.
+		func() {
+			defer a.gate.release(sessionID)
+			a.runChatTurn(pctx, ctx, sessionID, msg, func(agentEvent) {})
+		}()
+	}
+}
+
 func handleChat(w http.ResponseWriter, r *http.Request, a *agent, st *store) {
 	if r.Method != http.MethodPost {
 		http.Error(w, "method not allowed", 405)
@@ -227,6 +360,17 @@ func handleChat(w http.ResponseWriter, r *http.Request, a *agent, st *store) {
 	w.Header().Set("X-Accel-Buffering", "no") // disable proxy buffering
 	w.WriteHeader(200)
 
+	// All writes to w (events + the keepalive comment below) go through one
+	// mutex: http.ResponseWriter is NOT safe for concurrent use, and the
+	// keepalive ticker runs alongside the turn's event sink (plan 2026-08-03
+	// F3). Without this, interleaved writes corrupt the SSE stream.
+	var writeMu sync.Mutex
+	writeEvent := func(ev agentEvent) {
+		writeMu.Lock()
+		defer writeMu.Unlock()
+		sseEvent(w, flusher, ev)
+	}
+
 	ctx := r.Context()
 	sessionID := req.SessionID
 
@@ -278,134 +422,65 @@ func handleChat(w http.ResponseWriter, r *http.Request, a *agent, st *store) {
 		st.answerQuestion(pctx, sessionID, qid, req.Message)
 	}
 
-	sseEvent(w, flusher, agentEvent{Type: "session", Data: sessionID, SessionID: sessionID})
+	writeEvent(agentEvent{Type: "session", Data: sessionID, SessionID: sessionID})
 
-	// F1 (plan 2026-08-03): serialize turns per session. The user message is
-	// already persisted above, so even if we can't run this turn right now it
-	// isn't lost. Wait briefly for a finishing background turn (continuation /
-	// resume) so the common case is seamless; if one is still running after
-	// that, tell the operator to retry rather than spawning a second
-	// concurrent turn (the interleaving this gate exists to prevent). On
-	// success the permit is held until this handler returns (stream + post-
-	// processing done); background resumeSession callers skip while it's held.
+	// F1/F2 (plan 2026-08-03): serialize turns per session. The user message is
+	// already persisted above, so it is never lost. Wait briefly for a finishing
+	// background turn; if one is still running after that, QUEUE this message
+	// (don't reject it) and tell the client so it shows a "queued" state. The
+	// in-flight turn's release drains the queue (drainQueued) and runs it as a
+	// real turn server-side. This never stacks concurrent turns — the gate still
+	// guarantees one in-flight turn per session.
 	const turnWait = 5 * time.Second
 	if !a.gate.acquire(sessionID, turnWait) {
-		slog.Info("nomos: turn already active, deferring operator message", "session", sessionID)
-		sseEvent(w, flusher, agentEvent{
-			Type: "error",
-			Data: "Nomos is still finishing a previous step. Your message was saved — give it a moment to finish, then send it again.",
-		})
-		sseEvent(w, flusher, agentEvent{Type: "done", Data: map[string]any{
+		a.queue.enqueue(sessionID, req.Message)
+		slog.Info("nomos: turn already active, queued operator message", "session", sessionID)
+		writeEvent(agentEvent{Type: "queued", Data: sessionID, SessionID: sessionID})
+		writeEvent(agentEvent{Type: "done", Data: map[string]any{
 			"session_id": sessionID,
-			"error":      true,
+			"queued":     true,
 		}, SessionID: sessionID})
 		return
 	}
-	defer a.gate.release(sessionID)
+	defer func() {
+		a.gate.release(sessionID)
+		// Run any message that was queued while this turn held the gate. In a
+		// goroutine so the HTTP response finishes without waiting on the next
+		// turn; the queued turn has no SSE client of its own.
+		safego.Go("nomos:drain:"+sessionID, func() { a.drainQueued(context.Background(), sessionID) })
+	}()
 
-	toolCalls := []map[string]any{}
-	// P3: accumulate per-iteration reasoning instead of overwriting with
-	// the final `text` event. The agent loop emits a `text` event for each
-	// LLM iteration that produced text (intermediate reasoning before tool
-	// calls + the final answer). Without accumulation, only the last `text`
-	// survives in the persisted row — a reload shows the final summary but
-	// not the thinking that led to each tool call.
-	var textParts []string
-	var finalText string
-
-	// Incremental persistence, mirroring resumeSession's existing
-	// placeholder+update pattern (continue.go): insert a placeholder now,
-	// update the SAME row after every tool call, so whatever happened before
-	// an abort is never lost — only what hadn't happened yet is.
-	placeholder, _ := json.Marshal(map[string]any{"role": "assistant", "text": ""})
-	msgID, err := st.insertMessageReturningID(pctx, sessionID, "assistant", placeholder)
-	if err != nil {
-		slog.Error("nomos: chat placeholder insert failed", "session", sessionID, "error", err)
-	}
-	persist := func() {
-		if msgID == uuid.Nil {
-			return
-		}
-		body, _ := json.Marshal(map[string]any{
-			"role":       "assistant",
-			"text":       finalText,
-			"tool_calls": toolCalls,
-		})
-		st.updateMessage(pctx, msgID, body)
-	}
-
-	a.chat(ctx, sessionID, req.Message, func(ev agentEvent) {
-		if ev.Type == "tool_use" || ev.Type == "tool_result" {
-			if m, ok := ev.Data.(map[string]any); ok {
-				m["type"] = ev.Type
-				// One entry per tool call: tool_use creates it, tool_result
-				// merges the result into the same entry (matched by id).
-				// Before this fix, both events appended separate entries,
-				// doubling every tool call in the persisted transcript
-				// (confirmed pre-existing in d9cdcee1, v0.3.x era).
-				id, _ := m["id"].(string)
-				if id != "" && ev.Type == "tool_result" {
-					for _, existing := range toolCalls {
-						if eID, _ := existing["id"].(string); eID == id {
-							for k, v := range m {
-								existing[k] = v
-							}
-							break
-						}
-					}
-				} else {
-					toolCalls = append(toolCalls, m)
-				}
-			}
-			persist() // live: survives even if the client disconnects right after
-		}
-		if ev.Type == "text" {
-			// P3: accumulate. Each `text` event is one iteration's reasoning
-			// (or the final answer). Join with newlines so the persisted row
-			// reads as the full transcript of what the agent said, not just
-			// the last thing.
-			if t, ok := ev.Data.(string); ok && t != "" {
-				textParts = append(textParts, t)
-				finalText = strings.Join(textParts, "\n\n")
-				persist()
+	// F3 (plan 2026-08-03): keep the SSE alive during long turns. A turn can
+	// run for many minutes (provisioning chains, deep research); the model
+	// often takes 20-40s between tool iterations, and with nothing flushed in
+	// that gap a proxy/browser idle timeout silently closes the stream. The
+	// client then sees streaming=false while the server keeps working — the
+	// "I can't tell it's working" desync. An SSE comment line (":keepalive") is
+	// ignored by EventSource but resets idle timers.
+	keepDone := make(chan struct{})
+	go func() {
+		t := time.NewTicker(12 * time.Second)
+		defer t.Stop()
+		for {
+			select {
+			case <-keepDone:
+				return
+			case <-t.C:
+				writeMu.Lock()
+				fmt.Fprintf(w, ":keepalive\n\n")
+				flusher.Flush()
+				writeMu.Unlock()
 			}
 		}
-		sseEvent(w, flusher, ev)
+	}()
+	// Defer the close (not a statement after runChatTurn) so the goroutine
+	// exits even if runChatTurn panics — net/http recovers handler panics, so
+	// a non-deferred close would be skipped and the ticker would keep writing
+	// to a dead ResponseWriter forever.
+	defer close(keepDone)
+	a.runChatTurn(pctx, ctx, sessionID, req.Message, func(ev agentEvent) {
+		writeEvent(ev)
 	})
-
-	// B.6: if the turn ended with no text and no tool calls (the model
-	// empty-response'd and all retries failed), delete the placeholder row
-	// instead of persisting an empty bubble. The error event was already
-	// streamed to the frontend via the 'done with error=true' event, so the
-	// operator sees the error inline — an empty assistant bubble in the
-	// transcript adds nothing and looks like the agent is broken.
-	if finalText == "" && len(toolCalls) == 0 && msgID != uuid.Nil {
-		st.deleteMessage(pctx, msgID)
-	} else {
-		persist() // final state — same row, updated one last time with the concluding text
-	}
-
-	// Generate a meaningful title from the assistant's first answer
-	// instead of reusing the raw user message for every session.
-	// P2.9 (2026-07-20): prefer the goal as the title when one is set —
-	// the first assistant text is often a greeting or narrative that
-	// doesn't describe the task ("Hey! 👋 Nomos here, running on
-	// mac-mini:8092..."). The goal is the operator's actual intent.
-	// Sessions that never call set_goal (pure Q&A) fall back to the
-	// assistant text, which is still better than the raw user message.
-	if finalText != "" && sessionID != "ephemeral" {
-		var goalTitle string
-		if sess, gerr := st.getSession(pctx, sessionID); gerr == nil && sess.Goal != "" {
-			goalTitle = truncate(sess.Goal, 120)
-		}
-		title := goalTitle
-		if title == "" {
-			title = truncate(finalText, 80)
-		}
-		if title != "" {
-			st.updateSessionTitle(pctx, sessionID, title)
-		}
-	}
 }
 
 func handleSessionsList(w http.ResponseWriter, r *http.Request, st *store) {

cmd/nomos/messagequeue.go

@@ -0,0 +1,82 @@
+package main
+
+import (
+	"log/slog"
+	"sync"
+)
+
+// maxQueuedPerSession caps a session's queue. A held turn plus unbounded
+// enqueues would grow memory without limit; an operator nudging a long
+// autonomous turn realistically queues only a handful, so a generous cap is
+// pure insurance. Overflow drops the newest enqueue and logs (the message is
+// already persisted in the DB by handleChat before enqueue, so it isn't lost
+// from the transcript — it just won't auto-run).
+const maxQueuedPerSession = 20
+
+// messageQueue holds operator messages that arrived while a turn was already
+// running for a session. Plan 2026-08-03 (F2): instead of rejecting the
+// operator's message with "Nomos is still finishing a previous step… send it
+// again", the message is queued and auto-run when the in-flight turn releases
+// the session's turn-gate permit.
+//
+// The queue only schedules WHEN a turn runs, not WHETHER the message is stored
+// — handleChat persists the user message before acquiring the gate, so a queued
+// message is already in the transcript; this just makes sure a turn eventually
+// acts on it.
+//
+// Draining is strictly one-at-a-time under the turn gate (see drainQueued in
+// main.go), so this cannot stack concurrent turns — the exact hazard the gate
+// itself exists to prevent. Background resumeSession callers never touch this
+// queue; they keep their non-blocking skip.
+type messageQueue struct {
+	mu    sync.Mutex
+	queue map[string][]string
+}
+
+func newMessageQueue() *messageQueue {
+	return &messageQueue{queue: map[string][]string{}}
+}
+
+// enqueue appends a message to the back of the session's FIFO. Returns false
+// (and logs) if the session is already at maxQueuedPerSession — the caller's
+// message is already persisted in the DB, so this only skips auto-running it.
+func (q *messageQueue) enqueue(sessionID, msg string) bool {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	if len(q.queue[sessionID]) >= maxQueuedPerSession {
+		slog.Warn("nomos: message queue full; dropping auto-run for operator message", "session", sessionID, "cap", maxQueuedPerSession)
+		return false
+	}
+	q.queue[sessionID] = append(q.queue[sessionID], msg)
+	return true
+}
+
+// dequeue pops the next message from the front of the session's FIFO. Returns
+// ok=false when empty.
+func (q *messageQueue) dequeue(sessionID string) (string, bool) {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	xs := q.queue[sessionID]
+	if len(xs) == 0 {
+		return "", false
+	}
+	m := xs[0]
+	q.queue[sessionID] = xs[1:]
+	return m, true
+}
+
+// requeueFront pushes a message back to the front — used when a drainer popped
+// a message but lost the race for the gate to a live turn; that turn's own
+// release will drain it again.
+func (q *messageQueue) requeueFront(sessionID, msg string) {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	q.queue[sessionID] = append([]string{msg}, q.queue[sessionID]...)
+}
+
+// peek reports the queued depth for a session (test/diagnostic helper).
+func (q *messageQueue) peek(sessionID string) int {
+	q.mu.Lock()
+	defer q.mu.Unlock()
+	return len(q.queue[sessionID])
+}

cmd/nomos/messagequeue_test.go

@@ -0,0 +1,142 @@
+package main
+
+import (
+	"context"
+	"sync"
+	"testing"
+	"time"
+)
+
+func TestMessageQueue_FIFO(t *testing.T) {
+	q := newMessageQueue()
+	q.enqueue("s", "first")
+	q.enqueue("s", "second")
+	q.enqueue("s", "third")
+
+	want := []string{"first", "second", "third"}
+	for _, w := range want {
+		got, ok := q.dequeue("s")
+		if !ok || got != w {
+			t.Fatalf("dequeue = %q,%v want %q,true", got, ok, w)
+		}
+	}
+	if _, ok := q.dequeue("s"); ok {
+		t.Fatal("dequeue on drained queue should return ok=false")
+	}
+}
+
+func TestMessageQueue_RequeueFront(t *testing.T) {
+	q := newMessageQueue()
+	q.enqueue("s", "a")
+	q.enqueue("s", "b")
+	// Pop "a", then push it back to the front; "a" must come out before "b".
+	a, _ := q.dequeue("s")
+	q.requeueFront("s", a)
+	got, _ := q.dequeue("s")
+	if got != "a" {
+		t.Fatalf("after requeueFront, dequeue = %q want %q", got, "a")
+	}
+	got2, _ := q.dequeue("s")
+	if got2 != "b" {
+		t.Fatalf("next dequeue = %q want %q", got2, "b")
+	}
+}
+
+func TestMessageQueue_IsolatedPerSession(t *testing.T) {
+	q := newMessageQueue()
+	q.enqueue("s1", "one")
+	q.enqueue("s2", "two")
+	if got, _ := q.dequeue("s1"); got != "one" {
+		t.Fatalf("s1 = %q want one", got)
+	}
+	if got, _ := q.dequeue("s2"); got != "two" {
+		t.Fatalf("s2 = %q want two", got)
+	}
+	if q.peek("s1") != 0 || q.peek("s2") != 0 {
+		t.Fatal("both sessions should be drained")
+	}
+}
+
+func TestMessageQueue_Concurrent(t *testing.T) {
+	q := newMessageQueue()
+	const n = maxQueuedPerSession // stay under the cap so every enqueue lands
+	var wg sync.WaitGroup
+	for i := 0; i < n; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			q.enqueue("s", "m")
+		}(i)
+	}
+	wg.Wait()
+	if q.peek("s") != n {
+		t.Fatalf("peek = %d want %d (all enqueues must be counted)", q.peek("s"), n)
+	}
+	seen := 0
+	for {
+		if _, ok := q.dequeue("s"); !ok {
+			break
+		}
+		seen++
+	}
+	if seen != n {
+		t.Fatalf("drained %d want %d", seen, n)
+	}
+}
+
+func TestMessageQueue_CapsOverflow(t *testing.T) {
+	q := newMessageQueue()
+	for i := 0; i < maxQueuedPerSession; i++ {
+		if !q.enqueue("s", "m") {
+			t.Fatalf("enqueue #%d within cap should succeed", i)
+		}
+	}
+	if q.enqueue("s", "overflow") {
+		t.Fatal("enqueue past the cap should return false (dropped)")
+	}
+	if got := q.peek("s"); got != maxQueuedPerSession {
+		t.Fatalf("peek = %d want %d (overflow must not append)", got, maxQueuedPerSession)
+	}
+}
+
+// drainQueued on an empty queue must be a no-op: it returns immediately and
+// never touches the gate (so the session stays free for the next turn).
+func TestDrainQueued_NoOpOnEmpty(t *testing.T) {
+	a := &agent{gate: newTurnGate(), queue: newMessageQueue()}
+	a.drainQueued(context.Background(), "s")
+	if !a.gate.acquire("s", 0) {
+		t.Fatal("gate should be free after a no-op drain (drain must not hold it)")
+	}
+	a.gate.release("s")
+}
+
+// With a queued message but the gate held by another turn, drainQueued must
+// re-queue the message and return WITHOUT running a turn (no store/provider → a
+// real run would panic). This is the "never stack" property: a busy gate
+// defers to the holder's own release-drain.
+func TestDrainQueued_RequeuesWhenBusy(t *testing.T) {
+	prev := drainAcquireWait
+	drainAcquireWait = 10 * time.Millisecond
+	t.Cleanup(func() { drainAcquireWait = prev })
+
+	a := &agent{gate: newTurnGate(), queue: newMessageQueue()}
+	if !a.gate.acquire("s", 0) {
+		t.Fatal("precondition: hold the gate")
+	}
+	a.queue.enqueue("s", "queued-msg")
+
+	done := make(chan struct{})
+	go func() {
+		a.drainQueued(context.Background(), "s") // must not panic; must requeue
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(time.Second):
+		t.Fatal("drainQueued did not return promptly while the gate was busy")
+	}
+	if got := a.queue.peek("s"); got != 1 {
+		t.Fatalf("message should be re-queued while busy; peek = %d want 1", got)
+	}
+	a.gate.release("s")
+}