fix(selfhost): deliver MULTICA_LLM_* to the backend and log the layer state

SELF_HOSTING_AI.md

@@ -80,6 +80,29 @@ invoked directly is the reverse again — there the environment outranks `.env`.
 startup output is read back from Docker Compose, so the address it prints is the
 one the stack is actually published on.
 
+## Optional: server-side LLM features
+
+Chat auto-titling and chat follow-up suggestions ("quick actions") are produced
+by a small server-side model call, separate from the agent runtimes. They are
+off until the backend has an upstream configured:
+
+```bash
+# In .env — either the key or the base URL is enough to enable the layer.
+MULTICA_LLM_API_KEY=sk-...
+MULTICA_LLM_BASE_URL=          # optional: any OpenAI-compatible gateway
+MULTICA_LLM_DEFAULT_MODEL=     # optional: defaults to a small built-in model
+```
+
+Both features degrade silently when this is unset — titles stay as the first
+message and no follow-up suggestions appear — so confirm the state from the
+backend's startup log rather than from the UI:
+
+```bash
+docker compose -f docker-compose.selfhost.yml logs backend | grep "llm layer"
+# "llm layer enabled"  -> configured
+# "llm layer disabled" -> unset; the two features above will not appear
+```
+
 ## Troubleshooting
 
 - **Backend not ready:** `docker compose -f docker-compose.selfhost.yml logs backend`

deploy/helm/multica/values.yaml

@@ -43,6 +43,13 @@ images:
 # The chart references this Secret by name; it does not template it, so real
 # values never need to land in git.
 # -----------------------------------------------------------------------------
+# Name of a pre-created Secret mounted onto the backend with envFrom, so any
+# key in it becomes an environment variable. Server-side LLM features (chat
+# auto-titling, chat follow-up suggestions) are configured here rather than in
+# this file, because they are credentials: add MULTICA_LLM_API_KEY (and
+# optionally MULTICA_LLM_BASE_URL / MULTICA_LLM_DEFAULT_MODEL) to this Secret.
+# Without them the backend logs "llm layer disabled" at startup and both
+# features stay silently off.
 existingSecret: multica-secrets
 
 # -----------------------------------------------------------------------------

docker-compose.selfhost.yml

@@ -102,6 +102,14 @@ services:
       DISABLE_WORKSPACE_CREATION: ${DISABLE_WORKSPACE_CREATION:-}
       GITHUB_APP_SLUG: ${GITHUB_APP_SLUG:-}
       GITHUB_WEBHOOK_SECRET: ${GITHUB_WEBHOOK_SECRET:-}
+      # Server-internal LLM layer (MUL-4238). Backs chat auto-titling and chat
+      # follow-up suggestions. Setting either the key or the base URL enables
+      # it; leaving both empty disables those features silently. Without these
+      # three lines the vars in .env never reach the container, so the features
+      # cannot be turned on at all no matter what the operator configures.
+      MULTICA_LLM_API_KEY: ${MULTICA_LLM_API_KEY:-}
+      MULTICA_LLM_BASE_URL: ${MULTICA_LLM_BASE_URL:-}
+      MULTICA_LLM_DEFAULT_MODEL: ${MULTICA_LLM_DEFAULT_MODEL:-}
       # Public URL the API is reachable at from the open internet, no
       # trailing slash. Used to mint absolute webhook URLs for autopilot
       # webhook triggers. Leave unset behind a same-origin reverse proxy

server/internal/handler/chat_input_ownership_test.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/multica-ai/multica/server/internal/service"
 	db "github.com/multica-ai/multica/server/pkg/db/generated"
@@ -674,6 +675,48 @@ func TestChatQuickActions_ContextAnchorsOnTargetTurn(t *testing.T) {
 	}
 }
 
+// TestChatQuickActions_AutomaticPassEndToEnd exercises the path a real chat
+// turn takes — CompleteTask decides eligibility, broadcasts, and starts the
+// pass — rather than calling the generator directly. Everything between the
+// completion callback and the pills landing on the row is only covered here.
+func TestChatQuickActions_AutomaticPassEndToEnd(t *testing.T) {
+	if testHandler == nil {
+		t.Skip("database not available")
+	}
+	ctx := context.Background()
+	agentID, sessionID, _, _ := setupDirectChatSession(t, ctx, "quick-actions auto e2e")
+
+	taskID := sendDirectChat(t, ctx, agentID, sessionID, "what should I do next?")
+	markTaskRunning(t, ctx, taskID)
+
+	restore := installStubQuickActions()
+	defer restore()
+
+	if _, err := testHandler.TaskService.CompleteTask(
+		ctx, parseUUID(taskID), completeResult(t, "Here is the plan."), "", "", false, ""); err != nil {
+		t.Fatalf("complete task: %v", err)
+	}
+
+	// The pass runs on a detached goroutine; poll for its write.
+	var actions []protocol.ChatQuickAction
+	for i := 0; i < 100; i++ {
+		rows := assistantRows(t, ctx, sessionID)
+		if len(rows) == 1 {
+			actions = nil
+			if err := json.Unmarshal(rows[0].QuickActions, &actions); err != nil {
+				t.Fatalf("decode quick actions: %v", err)
+			}
+			if len(actions) > 0 {
+				break
+			}
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	if len(actions) != 1 || actions[0].Label != "Next" {
+		t.Fatalf("automatic pass must attach suggestions to the completed turn, got %+v", actions)
+	}
+}
+
 // installStubQuickActions enables suggestion generation and returns the undo.
 // Scoped tightly around the synchronous RegenerateChatQuickActions calls rather
 // than installed for a whole test: CompleteTask starts a background pass when

server/internal/handler/handler.go

@@ -319,6 +319,17 @@ func New(queries *db.Queries, txStarter txStarter, hub *realtime.Hub, bus *event
 		DefaultModel: cfg.LLMDefaultModel,
 	})
 
+	// One boot-time line so an operator can tell configured-and-working apart
+	// from silently-off. Every consumer of this layer degrades quietly by
+	// design (auto-titling keeps the original title, chat follow-up suggestions
+	// simply never appear), which without this log makes an unset key
+	// indistinguishable from a broken feature.
+	if llmClient.Enabled() {
+		slog.Info("llm layer enabled", "default_model", llmClient.DefaultModel())
+	} else {
+		slog.Warn("llm layer disabled: set MULTICA_LLM_API_KEY or MULTICA_LLM_BASE_URL to enable chat auto-titling and chat follow-up suggestions")
+	}
+
 	taskSvc := service.NewTaskService(queries, txStarter, hub, bus, daemonHub)
 	taskSvc.Analytics = analyticsClient
 	// Chat follow-up suggestions run through the same internal LLM layer that