fix: chat session reliability, cost, and hygiene (empty-response guard, tool truncation, delete, titles)

- Empty/refusal responses retried once, then surfaced as errors instead of silent blanks
- Chinese refusal boilerplate detected via denylist + non-ASCII heuristic
- Bulk-tool preference added to SOUL.md (list_lxcs over per-entity get_lxc_state)
- Tool results truncated to 4KB on persist; get_state_snapshot filters null-state entities
- Session delete (DELETE /sessions/{id} + confirm-on-second-click UI)
- Session titles auto-generated from assistant answer instead of raw user message

cmd/nomos/agent.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"log/slog"
 	"os"
+	"strings"
 	"time"
 
 	"github.com/google/uuid"
@@ -15,6 +16,15 @@ import (
 )
 
 const maxIterations = 15
+const maxLLMRetries = 1
+
+var refusalDenylist = []string{
+	"我没有相关信息",
+	"您可以尝试问我其它问题",
+	"我无法",
+	"抱歉，我无法",
+	"关于这个问题，我没有",
+}
 
 type agent struct {
 	client   *mcpClient
@@ -105,14 +115,6 @@ func (a *agent) chat(ctx context.Context, sessionID, message string, emit func(a
 		return
 	}
 
-	// Rebuild conversation context from persisted history so sessions are
-	// multi-turn. The current user turn is saved by the HTTP handler before
-	// this runs, so it is already included in the history for real sessions.
-	// Prior tool_use/tool_result pairs are replayed as a tool-calling
-	// assistant message followed by matching tool-role results, so the agent
-	// starts each turn already knowing what it already checked instead of
-	// re-querying the same tools from scratch. Ephemeral sessions (no store)
-	// fall back to the single incoming message.
 	system := a.system
 	if snapshot := a.fleetSnapshot(); snapshot != "" {
 		system += "\n\n" + snapshot
@@ -147,30 +149,54 @@ func (a *agent) chat(ctx context.Context, sessionID, message string, emit func(a
 			Tools:    tools,
 		}
 
-		// Stream the completion, emitting token deltas as they arrive. The
-		// accumulator reassembles the full message (content + tool calls) for
-		// the loop's control flow.
-		stream := a.provider.Chat.Completions.NewStreaming(ctx, params, a.reqOpts...)
-		acc := openai.ChatCompletionAccumulator{}
-		for stream.Next() {
-			chunk := stream.Current()
-			acc.AddChunk(chunk)
-			if len(chunk.Choices) > 0 {
-				if delta := chunk.Choices[0].Delta.Content; delta != "" {
-					emit(agentEvent{Type: "text_delta", Data: delta, SessionID: sessionID, Iteration: i + 1})
+		var msg openai.ChatCompletionMessage
+		var acc openai.ChatCompletionAccumulator
+
+		for attempt := 0; attempt <= maxLLMRetries; attempt++ {
+			acc = openai.ChatCompletionAccumulator{}
+			stream := a.provider.Chat.Completions.NewStreaming(ctx, params, a.reqOpts...)
+			for stream.Next() {
+				chunk := stream.Current()
+				acc.AddChunk(chunk)
+				if len(chunk.Choices) > 0 {
+					if delta := chunk.Choices[0].Delta.Content; delta != "" {
+						emit(agentEvent{Type: "text_delta", Data: delta, SessionID: sessionID, Iteration: i + 1})
+					}
 				}
 			}
-		}
-		if err := stream.Err(); err != nil {
-			emit(agentEvent{Type: "error", Data: fmt.Sprintf("llm: %v", err), SessionID: sessionID})
-			return
-		}
-		if len(acc.Choices) == 0 {
-			emit(agentEvent{Type: "error", Data: "no choices in response", SessionID: sessionID})
-			return
-		}
+			if err := stream.Err(); err != nil {
+				if attempt < maxLLMRetries {
+					slog.Warn("nomos: llm stream error, retrying", "error", err, "attempt", attempt+1, "session", sessionID)
+					continue
+				}
+				emit(agentEvent{Type: "error", Data: fmt.Sprintf("llm: %v", err), SessionID: sessionID})
+				return
+			}
+			if len(acc.Choices) == 0 {
+				if attempt < maxLLMRetries {
+					slog.Warn("nomos: no choices in response, retrying", "attempt", attempt+1, "session", sessionID)
+					continue
+				}
+				emit(agentEvent{Type: "error", Data: "no choices in response", SessionID: sessionID})
+				return
+			}
 
-		msg := acc.Choices[0].Message
+			msg = acc.Choices[0].Message
+
+			if len(msg.ToolCalls) == 0 {
+				if isRefusalOrEmpty(msg.Content) {
+					if attempt < maxLLMRetries {
+						slog.Warn("nomos: empty or refusal response, retrying",
+							"session", sessionID, "iter", i+1, "attempt", attempt+1,
+							"content_len", len(msg.Content))
+						continue
+					}
+					emit(agentEvent{Type: "error", Data: "Nomos returned an empty or unusable response — please retry.", SessionID: sessionID})
+					return
+				}
+			}
+			break
+		}
 
 		if len(msg.ToolCalls) == 0 {
 			emit(agentEvent{Type: "text", Data: msg.Content, SessionID: sessionID})
@@ -388,6 +414,33 @@ func (a *agent) fleetSnapshot() string {
 	return summary
 }
 
+// isRefusalOrEmpty returns true when the LLM response is blank or looks like a
+// canned non-English refusal to an English-language conversation. Flash-tier
+// models occasionally emit Chinese boilerplate deflection instead of a real
+// answer; this catches it before it reaches the UI.
+func isRefusalOrEmpty(text string) bool {
+	if strings.TrimSpace(text) == "" {
+		return true
+	}
+	ascii, nonASCII := 0, 0
+	for _, r := range text {
+		if r <= 127 {
+			ascii++
+		} else {
+			nonASCII++
+		}
+	}
+	if nonASCII > ascii {
+		return true
+	}
+	for _, pattern := range refusalDenylist {
+		if strings.Contains(text, pattern) {
+			return true
+		}
+	}
+	return false
+}
+
 func (a *agent) buildTools() ([]openai.ChatCompletionToolParam, error) {
 	defs, err := a.client.listToolsFull()
 	if err != nil {

cmd/nomos/main.go

@@ -186,6 +186,15 @@ func handleChat(w http.ResponseWriter, r *http.Request, a *agent, st *store) {
 		"tool_calls": toolCalls,
 	})
 	st.saveMessage(ctx, sessionID, "assistant", assistantMsg)
+
+	// Generate a meaningful title from the assistant's first answer
+	// instead of reusing the raw user message for every session.
+	if finalText != "" && sessionID != "ephemeral" {
+		title := truncate(finalText, 80)
+		if title != "" {
+			st.updateSessionTitle(ctx, sessionID, title)
+		}
+	}
 }
 
 func handleSessionsList(w http.ResponseWriter, r *http.Request, st *store) {
@@ -220,13 +229,26 @@ func handleSessionDetail(w http.ResponseWriter, r *http.Request, st *store) {
 		return
 	}
 
-	messages, err := st.getMessages(r.Context(), id)
-	if err != nil {
-		http.Error(w, err.Error(), 500)
-		return
+	switch r.Method {
+	case http.MethodDelete:
+		if err := st.deleteSession(r.Context(), id); err != nil {
+			http.Error(w, err.Error(), 500)
+			return
+		}
+		w.WriteHeader(204)
+
+	case http.MethodGet:
+		messages, err := st.getMessages(r.Context(), id)
+		if err != nil {
+			http.Error(w, err.Error(), 500)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]any{"session_id": id, "messages": messages})
+
+	default:
+		http.Error(w, "method not allowed", 405)
 	}
-	w.Header().Set("Content-Type", "application/json")
-	json.NewEncoder(w).Encode(map[string]any{"session_id": id, "messages": messages})
 }
 
 func handleQuery(w http.ResponseWriter, r *http.Request, client *mcpClient, agentSlug, mcpURL string) {

cmd/nomos/store.go

@@ -10,6 +10,8 @@ import (
 	"github.com/jackc/pgx/v5/pgxpool"
 )
 
+const maxToolResultSize = 4096
+
 type store struct {
 	pool *pgxpool.Pool
 }
@@ -71,10 +73,45 @@ func (s *store) saveMessage(ctx context.Context, sessionID, role string, content
 	}
 	_, err := s.pool.Exec(ctx,
 		`INSERT INTO agent_messages (session_id, role, content) VALUES ($1, $2, $3)`,
-		sessionID, role, content)
+		sessionID, role, truncateToolResults(content))
 	return err
 }
 
+func truncateToolResults(content json.RawMessage) json.RawMessage {
+	var m map[string]any
+	if err := json.Unmarshal(content, &m); err != nil {
+		return content
+	}
+	toolCalls, ok := m["tool_calls"].([]any)
+	if !ok || len(toolCalls) == 0 {
+		return content
+	}
+	changed := false
+	for i, raw := range toolCalls {
+		tc, ok := raw.(map[string]any)
+		if !ok {
+			continue
+		}
+		if result, ok := tc["result"]; ok {
+			resultJSON, _ := json.Marshal(result)
+			if len(resultJSON) > maxToolResultSize {
+				tc["result"] = string(resultJSON[:maxToolResultSize]) + fmt.Sprintf("...truncated (%d bytes total)", len(resultJSON))
+				toolCalls[i] = tc
+				changed = true
+			}
+		}
+	}
+	if !changed {
+		return content
+	}
+	m["tool_calls"] = toolCalls
+	out, err := json.Marshal(m)
+	if err != nil {
+		return content
+	}
+	return out
+}
+
 func (s *store) touchSession(ctx context.Context, id string) {
 	if s != nil {
 		s.pool.Exec(ctx, `UPDATE agent_sessions SET last_active_at=now() WHERE id=$1`, id)
@@ -126,6 +163,26 @@ func (s *store) getMessages(ctx context.Context, sessionID string) ([]message, e
 	return out, rows.Err()
 }
 
+func (s *store) deleteSession(ctx context.Context, id string) error {
+	if s == nil {
+		return nil
+	}
+	_, err := s.pool.Exec(ctx, `DELETE FROM agent_messages WHERE session_id = $1`, id)
+	if err != nil {
+		return err
+	}
+	_, err = s.pool.Exec(ctx, `DELETE FROM agent_sessions WHERE id = $1`, id)
+	return err
+}
+
+func (s *store) updateSessionTitle(ctx context.Context, id, title string) error {
+	if s == nil {
+		return nil
+	}
+	_, err := s.pool.Exec(ctx, `UPDATE agent_sessions SET title = $1 WHERE id = $2`, title, id)
+	return err
+}
+
 // resolveAgentID looks up the UUID of the agent entity (e.g. "agent:nomos").
 // Returns uuid.Nil if the store is absent or the slug is unknown.
 func (s *store) resolveAgentID(ctx context.Context, slug string) uuid.UUID {