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
2026-07-09 10:18:06 +02:00
parent 614c38ea7c
commit 49c37fe8b1
9 changed files with 335 additions and 60 deletions
--- a/cmd/nomos/agent.go
+++ b/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 {
--- a/cmd/nomos/main.go
+++ b/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) {
--- a/cmd/nomos/store.go
+++ b/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 {
--- a/internal/mcp/server.go
+++ b/internal/mcp/server.go
@@ -698,6 +698,8 @@ func newServer(pool *db.Pool, agentID uuid.UUID) *mcp.Server {
 			       COALESCE(st.last_check_at::text, '') AS last_check
 			FROM entities e
 			LEFT JOIN entity_status st ON st.entity_id = e.id
+			WHERE e.state IS NOT NULL
+			   OR st.health IS NOT NULL
 			ORDER BY st.health, e.slug
 			LIMIT 200
 		`), nil
--- a/nomos/SOUL.md
+++ b/nomos/SOUL.md
@@ -23,14 +23,30 @@ the actuator (a separate container with restricted SSH key) picks up.

 ## Key MCP tools

- `get_entity`, `list_entities` — resolve slugs to state
+- `list_lxcs` — all LXC containers with host, IP, health (use for fleet-wide questions)
+- `get_lxc_state` — per-container `pct status` (use only for a specific named container)
+- `get_state_snapshot` — fleet health, disk, drift at a glance
+- `get_health_summary` — fleet health counts
+- `query_metrics` — time-series metrics (prefer over per-entity `get_trend` for fleet-wide)
+- `list_entities` — resolve slugs to state (pass `type` filter when possible)
+- `get_entity` — single-entity detail
 - `get_blast_radius` — understand impact before requesting action
- `get_health_summary` — fleet status at a glance
 - `get_signal_history` — open alerts
- `get_trend` — metric trends for decisions
+- `get_trend` — metric trends for a specific entity (single-entity only)
 - `request_execution` — the ONLY mutation path
 - `get_agent_activity` — your own behavior log

+### Tool selection rules
+
+- **Fleet-wide questions** (e.g. "which hosts are saturated?", "what needs updating?"):
+  prefer bulk tools: `list_lxcs`, `get_health_summary`, `get_state_snapshot`,
+  `query_metrics`. Only fall back to per-entity tools (`get_lxc_state`, `tail_log`,
+  `get_trend`) for a specific named entity the user asked about.
+- **One call > many calls**: each `get_lxc_state` is a live SSH round-trip.
+  `list_lxcs` answers the same question in one call. Use it.
+- When a bulk tool's summary isn't enough for a specific entity, call the
+  per-entity tool for that one entity — not for every entity in the fleet.
+
 ## Policy awareness

 Before calling `request_execution`:
--- a/web/src/lib/api.ts
+++ b/web/src/lib/api.ts
@@ -31,6 +31,11 @@ export async function fetchMessages(sessionId: string): Promise<Message[]> {
  return data.messages ?? []
 }

+export async function deleteSession(sessionId: string): Promise<boolean> {
+  const res = await fetch(`${BASE}/sessions/${sessionId}`, { method: 'DELETE' })
+  return res.ok
+}
+
 export interface ChatEvent {
  type: string
  data: any
--- a/web/src/lib/components/SessionRail.svelte
+++ b/web/src/lib/components/SessionRail.svelte
@@ -1,20 +1,39 @@
 <script lang="ts">
  import { onMount } from 'svelte'
-  import { sessions, currentSession, loadSessions, loadSessionMessages, newChat } from '$lib/stores/chat'
+  import { sessions, currentSession, loadSessions, loadSessionMessages, newChat, deleteSession } from '$lib/stores/chat'
  import { relativeTime } from '$lib/utils'
  import { Button } from '$lib/components/ui/button'
  import { ScrollArea } from '$lib/components/ui/scroll-area'
  import PlusIcon from '@lucide/svelte/icons/plus'
+  import Trash2Icon from '@lucide/svelte/icons/trash2'

  onMount(() => {
    loadSessions()
  })

-  // Pick up sessions created/renamed elsewhere (e.g. after a turn completes).
  $effect(() => {
    void $currentSession
    loadSessions()
  })
+
+  let confirmDelete = $state<string | null>(null)
+
+  function handleDelete(e: MouseEvent, id: string) {
+    e.stopPropagation()
+    if (confirmDelete === id) {
+      deleteSession(id)
+      confirmDelete = null
+    } else {
+      confirmDelete = id
+      // Hide confirmation after 3s
+      setTimeout(() => { if (confirmDelete === id) confirmDelete = null }, 3000)
+    }
+  }
+
+  function handleClick(sessionId: string) {
+    confirmDelete = null
+    loadSessionMessages(sessionId)
+  }
 </script>

 <aside class="flex h-full w-56 shrink-0 flex-col gap-2 overflow-y-auto border-r bg-card/50 p-2">
@@ -27,10 +46,23 @@
      {#each $sessions as session (session.id)}
        <button
          type="button"
-          class="flex flex-col items-start gap-0.5 rounded-md border px-2 py-1.5 text-left text-xs transition-colors hover:bg-muted/60 {$currentSession === session.id ? 'border-primary bg-muted/50' : 'border-transparent'}"
-          onclick={() => loadSessionMessages(session.id)}
+          class="group flex flex-col items-start gap-0.5 rounded-md border px-2 py-1.5 text-left text-xs transition-colors hover:bg-muted/60 {$currentSession === session.id ? 'border-primary bg-muted/50' : 'border-transparent'}"
+          onclick={() => handleClick(session.id)}
        >
-          <span class="w-full truncate font-medium">{session.title || 'Untitled'}</span>
+          <span class="flex w-full items-center justify-between gap-1">
+            <span class="min-w-0 truncate font-medium">{session.title || 'Untitled'}</span>
+            <span
+              class="shrink-0 rounded p-0.5 opacity-0 transition-opacity group-hover:opacity-100 hover:bg-destructive/20 hover:text-destructive"
+              onclick={(e) => handleDelete(e, session.id)}
+              title={confirmDelete === session.id ? 'Click again to confirm delete' : 'Delete session'}
+            >
+              {#if confirmDelete === session.id}
+                <span class="text-[10px] font-semibold text-destructive">Sure?</span>
+              {:else}
+                <Trash2Icon class="size-3" />
+              {/if}
+            </span>
+          </span>
          <span class="text-[11px] text-muted-foreground">{relativeTime(session.last_active_at)}</span>
        </button>
      {:else}
--- a/web/src/lib/stores/chat.ts
+++ b/web/src/lib/stores/chat.ts
@@ -1,5 +1,5 @@
 import { writable, get } from 'svelte/store'
-import { streamChat, fetchSessions, fetchMessages } from '$lib/api'
+import { streamChat, fetchSessions, fetchMessages, deleteSession as apiDeleteSession } from '$lib/api'
 import type { ChatEvent, Session, Message } from '$lib/api'

 export interface ChatMessage {
@@ -177,3 +177,12 @@ export function cancelStream() {
    streaming.set(false)
  }
 }
+
+export async function deleteSession(sessionId: string) {
+  const ok = await apiDeleteSession(sessionId)
+  if (!ok) return
+  if (get(currentSession) === sessionId) {
+    newChat()
+  }
+  loadSessions()
+}
--- a/web/src/pages/Sessions.svelte
+++ b/web/src/pages/Sessions.svelte
@@ -1,29 +1,59 @@
 <script lang="ts">
-  import { sessions, loadSessions, loadSessionMessages, messages, currentSession } from '$lib/stores/chat'
+  import { sessions, loadSessions, loadSessionMessages, deleteSession, currentSession } from '$lib/stores/chat'
  import { onMount } from 'svelte'
+  import Trash2Icon from '@lucide/svelte/icons/trash2'

  onMount(() => {
    loadSessions()
  })
+
+  let confirmDelete = $state<string | null>(null)
+
+  function handleDelete(e: MouseEvent, id: string) {
+    e.stopPropagation()
+    if (confirmDelete === id) {
+      deleteSession(id)
+      confirmDelete = null
+    } else {
+      confirmDelete = id
+      setTimeout(() => { if (confirmDelete === id) confirmDelete = null }, 3000)
+    }
+  }
 </script>

 <div class="sessions-page">
  <h2>Sessions</h2>
  <div class="session-list">
    {#each $sessions as session (session.id)}
-      <button
-        class="session-card"
+      <div
+        class="session-card group"
        class:active={$currentSession === session.id}
-        onclick={() => {
-          loadSessionMessages(session.id)
-          location.hash = '#/chat'
-        }}
      >
-        <div class="session-title">{session.title || 'Untitled'}</div>
-        <div class="session-meta">
-          {new Date(session.last_active_at).toLocaleString()}
-        </div>
-      </button>
+        <button
+          class="session-content"
+          onclick={() => {
+            loadSessionMessages(session.id)
+            location.hash = '#/chat'
+          }}
+        >
+          <div class="session-title">{session.title || 'Untitled'}</div>
+          <div class="session-meta">
+            {new Date(session.last_active_at).toLocaleString()}
+          </div>
+        </button>
+        <button
+          class="session-delete"
+          onclick={(e) => handleDelete(e, session.id)}
+          title={confirmDelete === session.id ? 'Click again to confirm delete' : 'Delete session'}
+          aria-label="Delete session"
+        >
+          {#if confirmDelete === session.id}
+            <span class="confirm-text">Delete?</span>
+          {:else}
+            <Trash2Icon class="icon" />
+          {/if}
+        </button>
+      </div>
    {:else}
      <div class="empty">No sessions yet. Start chatting with Nomos.</div>
    {/each}
@@ -54,15 +84,12 @@
    display: flex;
    justify-content: space-between;
    align-items: center;
-    padding: 0.75rem 1rem;
    background: var(--bg-surface);
    border: 1px solid var(--border);
    border-radius: 8px;
-    cursor: pointer;
    color: var(--text);
    font-family: inherit;
    font-size: 0.875rem;
-    text-align: left;
    transition: border-color 0.15s;
  }

@@ -75,6 +102,58 @@
    background: var(--bg-hover);
  }

+  .session-content {
+    flex: 1;
+    display: flex;
+    justify-content: space-between;
+    align-items: center;
+    padding: 0.75rem 1rem;
+    background: none;
+    border: none;
+    cursor: pointer;
+    color: inherit;
+    font-family: inherit;
+    font-size: inherit;
+    text-align: left;
+  }
+
+  .session-delete {
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    width: 32px;
+    height: 32px;
+    margin-right: 4px;
+    padding: 0;
+    background: none;
+    border: none;
+    border-radius: 6px;
+    cursor: pointer;
+    color: var(--text-muted);
+    opacity: 0;
+    transition: opacity 0.15s, background 0.15s;
+  }
+
+  .group:hover .session-delete {
+    opacity: 1;
+  }
+
+  .session-delete:hover {
+    background: var(--destructive-bg-subtle, rgba(239, 68, 68, 0.1));
+    color: var(--destructive, #ef4444);
+  }
+
+  .confirm-text {
+    font-size: 0.65rem;
+    font-weight: 700;
+    color: var(--destructive, #ef4444);
+  }
+
+  .icon {
+    width: 16px;
+    height: 16px;
+  }
+
  .session-title {
    font-weight: 500;
  }