fix: LLM 工具循环 400、中断消息注入、ConPTY 终端支持

- agent: 工具轮请求尾部补 user 占位(zen 网关强制),tool 消息正确配对
- agent: 工具提醒/中断以 system 角色注入并带 [中断消息] 前缀,不进用户履历;系统提示词说明中断消息格式
- agentcli: 基于 ConPTY 的交互式终端(ptywin fork),terminal_create/read/write/resize/close/watch
- webui: server 输出通道适配器(保留 reasoning_content/disable_thinking)
- GUI: 沉浸式标题栏、icon 圆角重制、mascot 等打磨

internal/memory/document/document.go

@@ -183,6 +183,10 @@ func (s *Store) ContextToDoc(source string, entries []ContextEntry, vec vector.V
 		docVec = vec.Vectorize(summary + " " + content)
 	} else {
 		docVec = s.veczer.Vectorize(summary + " " + content)
+	}
+		meta := map[string]string{"content_hash": contentHash}
+	if source == "context_archived" {
+		meta["is_archived_context"] = "true"
 	}
 	doc := &Doc{
 		ID:          id,
@@ -195,7 +199,7 @@ func (s *Store) ContextToDoc(source string, entries []ContextEntry, vec vector.V
 		LastAccess:  time.Now(),
 		AccessCount: 1,
 		Source:      source,
-		Meta:        map[string]string{"content_hash": contentHash},
+		Meta:        meta,
 		Vector:      docVec,
 	}
 	s.docs[id] = doc

internal/memory/pipeline/pipeline.go

@@ -193,11 +193,15 @@ func (d *Distiller) distillLoop() {
 
 func (d *Distiller) distillOnce() {
 	d.mu.Lock()
-	cutoff := time.Now().AddDate(0, 0, -d.cfg.RetentionDays)
+	batchSize := d.cfg.BatchSize
+	if batchSize <= 0 {
+		batchSize = 50
+	}
+	// 每 tick 取前 N 条未蒸馏记录（无 RetentionDays 门槛），蒸馏成功才标记/移除
 	var toDistill []RawRecord
 	var remaining []RawRecord
 	for _, r := range d.records {
-		if r.CreatedAt.Before(cutoff) && !r.Distilled {
+		if !r.Distilled && len(toDistill) < batchSize {
 			toDistill = append(toDistill, r)
 		} else {
 			remaining = append(remaining, r)
@@ -210,22 +214,29 @@ func (d *Distiller) distillOnce() {
 		return
 	}
 
-	batchSize := d.cfg.BatchSize
-	if batchSize <= 0 {
-		batchSize = 50
-	}
+	distilled := 0
 	for i := 0; i < len(toDistill); i += batchSize {
 		end := i + batchSize
 		if end > len(toDistill) {
 			end = len(toDistill)
 		}
-		d.distillBatch(toDistill[i:end])
+		if d.distillBatch(toDistill[i:end]) {
+			distilled += end - i
+		} else {
+			// 蒸馏失败：记录写回待处理队列，下次 tick 重试
+			d.mu.Lock()
+			d.records = append(toDistill[i:end], d.records...)
+			d.mu.Unlock()
+		}
 	}
 	d.cleanupRawFiles()
-	log.Printf("[memory] distilled %d records", len(toDistill))
+	if distilled > 0 {
+		log.Printf("[memory] distilled %d records", distilled)
+	}
 }
 
-func (d *Distiller) distillBatch(batch []RawRecord) {
+// distillBatch 蒸馏一批记录，全部成功返回 true，任一失败返回 false（调用方重试）
+func (d *Distiller) distillBatch(batch []RawRecord) bool {
 	var userContent, assistantContent string
 	sessionIDs := make(map[string]bool)
 	for _, r := range batch {
@@ -245,8 +256,10 @@ func (d *Distiller) distillBatch(batch []RawRecord) {
 		}
 		if _, _, err := d.db.Commit(triples, sessionID, 0); err != nil {
 			log.Printf("[memory] distill commit: %v", err)
+			return false
 		}
 	}
+	return true
 }
 
 func (d *Distiller) cleanupRawFiles() {

internal/memory/pipeline/pipeline_test.go

@@ -1,6 +1,7 @@
 package pipeline
 
 import (
+	"fmt"
 	"os"
 	"path/filepath"
 	"testing"
@@ -117,6 +118,71 @@ func TestDistillOnce(t *testing.T) {
 	}
 }
 
+// Phase 4: 新记录无需等待 RetentionDays，下一 tick 立即蒸馏（文档所述 10min 频率）
+func TestDistillOnceFreshRecords(t *testing.T) {
+	db, err := memory.NewGraphDB(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer db.Close()
+
+	dir := t.TempDir()
+	d := NewDistiller(db, dir, DistillerConfig{
+		Interval:      10 * time.Minute,
+		RetentionDays: 7,
+		BatchSize:     50,
+	})
+	d.Append("sess1", "user", "我的名字是李四")
+	d.Append("sess1", "assistant", "你好李四！")
+
+	if len(d.records) != 2 {
+		t.Fatalf("expected 2 fresh records, got %d", len(d.records))
+	}
+
+	d.distillOnce()
+	if len(d.records) != 0 {
+		t.Errorf("fresh records should be distilled on next tick (no retention gate), got %d remaining", len(d.records))
+	}
+
+	// 二次蒸馏不重复（已蒸馏记录已被移除）
+	d.distillOnce()
+	if len(d.records) != 0 {
+		t.Errorf("second distill should be no-op, got %d records", len(d.records))
+	}
+}
+
+// Phase 4: BatchSize 限制每 tick 处理前 N 条，未蒸馏记录留待下个 tick
+func TestDistillOnceBatchLimit(t *testing.T) {
+	db, err := memory.NewGraphDB(filepath.Join(t.TempDir(), "test.db"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer db.Close()
+
+	dir := t.TempDir()
+	d := NewDistiller(db, dir, DistillerConfig{
+		Interval:      10 * time.Minute,
+		RetentionDays: 7,
+		BatchSize:     3,
+	})
+	for i := 0; i < 10; i++ {
+		d.Append("sess1", "user", fmt.Sprintf("第 %d 条消息内容", i))
+	}
+
+	d.distillOnce()
+	if len(d.records) != 7 {
+		t.Fatalf("expected 7 records remaining after batch 3, got %d", len(d.records))
+	}
+
+	// 后续 tick 继续消化，最终全部蒸馏
+	for i := 0; i < 5 && len(d.records) > 0; i++ {
+		d.distillOnce()
+	}
+	if len(d.records) != 0 {
+		t.Errorf("all records should be distilled after several ticks, got %d remaining", len(d.records))
+	}
+}
+
 func TestExtractKeyTriples(t *testing.T) {
 	tests := []struct {
 		user     string

internal/memory/static_embedder.go

@@ -129,12 +129,13 @@ func ensureModelFile(modelPath string) {
 	if modelPath == "" {
 		return
 	}
-	if _, err := os.Stat(modelPath); err == nil {
+	path, _ := parseModelSpec(modelPath)
+	if _, err := os.Stat(path); err == nil {
 		return
 	}
-	url := modelDownloadURL(modelPath)
-	log.Printf("[static_embedder] model %s not found, downloading from fastText...", modelPath)
-	if dlErr := downloadFastTextModel(modelPath, url); dlErr != nil {
+	url := modelDownloadURL(path)
+	log.Printf("[static_embedder] model %s not found, downloading from fastText...", path)
+	if dlErr := downloadFastTextModel(path, url); dlErr != nil {
 		log.Printf("[static_embedder] download failed: %v, will use TF-IDF fallback", dlErr)
 	} else {
 		log.Printf("[static_embedder] download ok")
@@ -183,7 +184,24 @@ func (e *StaticEmbedder) loadAll(paths []string) error {
 	return firstErr
 }
 
-func (e *StaticEmbedder) load(path string, primary bool) error {
+// parseModelSpec 解析模型路径规格：`path#top50000` 表示只加载前 50000 个词向量（按文件顺序，fastText
+// 词频降序，前 N 词覆盖绝大多数文本命中），用于降低常驻内存；无规格返回原路径与 0（全量加载）。
+func parseModelSpec(p string) (path string, topN int) {
+	path = p
+	if i := strings.IndexByte(p, '#'); i >= 0 {
+		path = p[:i]
+		spec := p[i+1:]
+		if strings.HasPrefix(spec, "top") {
+			if n, err := strconv.Atoi(strings.TrimPrefix(spec, "top")); err == nil && n > 0 {
+				topN = n
+			}
+		}
+	}
+	return path, topN
+}
+
+func (e *StaticEmbedder) load(spec string, primary bool) error {
+	path, topN := parseModelSpec(spec)
 	f, err := os.Open(path)
 	if err != nil {
 		return fmt.Errorf("open: %w", err)
@@ -217,7 +235,11 @@ func (e *StaticEmbedder) load(path string, primary bool) error {
 		vecSum = make([]float64, dim)
 	}
 
+	loaded := 0
 	for scanner.Scan() {
+		if topN > 0 && loaded >= topN {
+			break
+		}
 		line := strings.TrimSpace(scanner.Text())
 		if line == "" {
 			continue
@@ -244,6 +266,7 @@ func (e *StaticEmbedder) load(path string, primary bool) error {
 			}
 			count++
 		}
+		loaded++
 	}
 
 	if primary {
@@ -263,7 +286,7 @@ func (e *StaticEmbedder) load(path string, primary bool) error {
 		e.loaded = true
 	}
 
-	log.Printf("[static_embedder] loaded %d words, dim=%d from %s", len(e.words), e.dim, path)
+	log.Printf("[static_embedder] loaded %d words, dim=%d from %s (topN=%d)", len(e.words), e.dim, path, topN)
 	return nil
 }
 

internal/memory/synth_model_test.go

@@ -86,3 +86,45 @@ func newSynthEmbedder(t testing.TB, dim int) *StaticEmbedder {
 	}
 	return e
 }
+
+// Phase 5: #topN 规格裁剪加载——只加载前 N 个词向量，控制常驻内存
+func TestStaticEmbedderTopNSpec(t *testing.T) {
+	path := writeSynthModel(t, 300)
+
+	// 解析规格
+	cleanPath, topN := parseModelSpec(path + "#top5")
+	if cleanPath != path || topN != 5 {
+		t.Fatalf("parseModelSpec(#top5) = (%q, %d), want (%q, 5)", cleanPath, topN, path)
+	}
+	cleanPath2, topN2 := parseModelSpec(path)
+	if cleanPath2 != path || topN2 != 0 {
+		t.Fatalf("parseModelSpec(plain) = (%q, %d), want (%q, 0)", cleanPath2, topN2, path)
+	}
+	cleanPath3, topN3 := parseModelSpec(path + "#abc")
+	if cleanPath3 != path || topN3 != 0 {
+		t.Fatalf("parseModelSpec(#abc) = (%q, %d), want (%q, 0)", cleanPath3, topN3, path)
+	}
+
+	// 裁剪加载
+	e := NewStaticEmbedder(path + "#top5")
+	if !e.Loaded() {
+		t.Fatal("topN embedder should be loaded")
+	}
+	if len(e.words) != 5 {
+		t.Errorf("expected 5 words loaded with #top5, got %d", len(e.words))
+	}
+}
+
+// Phase 5: 裁剪后向量化仍可用（未命中词走 unkVec 兜底）
+func TestStaticEmbedderTopNVectorize(t *testing.T) {
+	path := writeSynthModel(t, 300)
+	e := NewStaticEmbedder(path + "#top1")
+	if !e.Loaded() {
+		t.Fatal("embedder should be loaded")
+	}
+	v := e.Vectorize("天气怎么样")
+	// 未命中词不应产生空向量（unkVec 兜底）
+	if len(v) == 0 {
+		t.Error("vectorize with topN=1 should still produce a vector (unkVec fallback)")
+	}
+}