diff --git a/app/search/tools.py b/app/search/tools.py
new file mode 100644
index 0000000..f9d8f7a
--- /dev/null
+++ b/app/search/tools.py
@@ -0,0 +1,321 @@
+"""搜索 Agent 工具调度器 — 工具描述与 dispatch 分发。
+
+实现 ``core/agent/protocols.ToolDispatcher`` Protocol。
+连接 TreeEnvironment（数据）、summarizer（LLM 摘要）、
+vision（VLM 观察）和 skills（策略加载）。
+
+与 TRM4 ``core/tree/tools.py`` 的差异:
+- 自由函数 ``dispatch()`` → ``SearchToolDispatcher`` 类（依赖注入）;
+- 同步 → 全异步;
+- view_node / search_similar 内部拆分为 env 数据读取 + summarizer LLM 摘要。
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+from app.search.summarizer import summarize_children, summarize_node, summarize_nodes_batch
+from app.search.vision import observe_frame
+from app.tree.environment import _LEVEL_LABEL, TreeEnvironment, _node_level
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+    from pathlib import Path
+
+    import numpy as np
+
+    from app.ports import OCRProvider
+    from app.search.skills import SkillRegistry
+    from core.protocols import LLMProvider, VLMProvider
+
+# ── 工具描述文本（与 TRM4 core/tree/tools.py 完全一致） ─────────────────
+
+_BASE_DESCRIPTIONS = """\
+## 可用工具
+
+在 action 中指定 tool 和 args 来调用工具。
+
+### view_node
+查看节点信息，获取与问题相关的内容摘要和子节点概览。
+- args: {"node_id": "节点 ID", "question": "当前关注的具体问题"}
+
+### search_similar
+语义检索最相关的节点，返回与问题相关的内容摘要。
+- args: {"query": "搜索关键词（2-4 词）", "question": "当前关注的具体问题", "k": 返回数量（可选，默认 5）}
+
+### observe_frame
+调用视觉模型查看关键帧图像，回答针对性的视觉问题。
+- args: {"node_ids": ["L3 节点 ID 列表（1-4 个），或单个 L2 节点 ID"], "question": "针对帧内容的具体视觉问题"}
+
+### submit_answer
+提交最终答案。
+- args: {"answer": "选项字母 A/B/C/D", "evidence": "关键证据摘要", "reasoning": "每个选项的判断理由"}"""
+
+_SKILL_DESCRIPTION = """
+
+### read_skill
+加载指定题型技能的详细搜索策略。
+- args: {"name": "技能名称"}"""
+
+
+def get_tool_descriptions(include_read_skill: bool = False) -> str:
+    """返回工具描述文本，用于写入 system prompt。
+
+    参数:
+        include_read_skill: 是否包含 read_skill 工具（manual 模式用）。
+
+    返回:
+        Markdown 格式的工具描述文本。
+    """
+    text = _BASE_DESCRIPTIONS
+    if include_read_skill:
+        text += _SKILL_DESCRIPTION
+    return text
+
+
+# ── SearchToolDispatcher ──────────────────────────────────────────────
+
+
+class SearchToolDispatcher:
+    """搜索 Agent 工具调度器，实现 ToolDispatcher Protocol。
+
+    按工具名路由到对应私有处理方法。未知工具抛 ValueError
+    （AgentLoop 捕获后不计步数）；节点不存在等运行时错误
+    捕获后返回错误文本。
+
+    参数:
+        env: 视频树运行时环境（纯数据访问）。
+        tool_llm: 摘要用 LLM 端口。
+        vlm: 视觉模型端口。
+        ocr: 帧文字转录端口（None 不启用）。
+        prompts_dir: prompt 文件目录。
+        skills: 技能注册表（None 不启用 read_skill）。
+        embed_fn: 文本嵌入函数（search_similar 用）。
+        verify_vision: observe_frame 是否执行验证轮。
+        anchor: view_node 是否启用行号锚模式。
+        assemble_mode: 锚模式装配形态（"ids"/"ids_expand"/"expand_only"）。
+        stats_sink: 统计回调（None 不收集）。
+    """
+
+    def __init__(
+        self,
+        env: TreeEnvironment,
+        tool_llm: LLMProvider,
+        vlm: VLMProvider,
+        ocr: OCRProvider | None,
+        prompts_dir: Path,
+        skills: SkillRegistry | None,
+        *,
+        embed_fn: Callable[[str | list[str]], np.ndarray],
+        verify_vision: bool,
+        anchor: bool,
+        assemble_mode: str,
+        stats_sink: Callable[[dict[str, Any]], None] | None = None,
+    ) -> None:
+        self._env = env
+        self._tool_llm = tool_llm
+        self._vlm = vlm
+        self._ocr = ocr
+        self._prompts_dir = prompts_dir
+        self._skills = skills
+        self._embed_fn = embed_fn
+        self._verify_vision = verify_vision
+        self._anchor = anchor
+        self._assemble_mode = assemble_mode
+        self._stats_sink = stats_sink
+
+    # ── ToolDispatcher Protocol 实现 ──────────────────────────────────
+
+    async def dispatch(
+        self, tool_name: str, args: dict[str, Any], *, context: dict[str, Any]
+    ) -> str:
+        """按工具名分发到对应处理方法。
+
+        参数:
+            tool_name: 工具名称。
+            args: 工具参数字典。
+            context: 调用上下文（含 session_id、parent_call_id 等遥测字段）。
+
+        返回:
+            工具执行结果文本。
+
+        异常:
+            ValueError: 未知工具名——上抛给 AgentLoop，不计步数。
+        """
+        try:
+            if tool_name == "view_node":
+                return await self._handle_view_node(args, context)
+            if tool_name == "search_similar":
+                return await self._handle_search_similar(args, context)
+            if tool_name == "observe_frame":
+                return await self._handle_observe_frame(args, context)
+            if tool_name == "submit_answer":
+                return f"[ok] 答案已提交: {args['answer']}"
+            if tool_name == "read_skill":
+                return self._handle_read_skill(args)
+        except (KeyError, FileNotFoundError) as e:
+            return f"工具执行错误: {e}"
+
+        raise ValueError(f"未知工具: {tool_name}")
+
+    # ── 私有处理方法 ──────────────────────────────────────────────────
+
+    async def _handle_view_node(self, args: dict[str, Any], context: dict[str, Any]) -> str:
+        """view_node：节点摘要 + 子节点概览。
+
+        参数:
+            args: {"node_id": str, "question": str}。
+            context: 遥测上下文。
+
+        返回:
+            "[节点] {id} | {level} | {time}\\n\\n{summary}\\n\\n[子节点概览] ..."
+        """
+        node_id: str = args["node_id"]
+        question: str = args["question"]
+        session_id = context.get("session_id")
+        parent_call_id = context.get("parent_call_id")
+
+        # Phase 1: 节点元数据（头部格式化）
+        node = self._env._id_to_node[node_id]
+        level = _node_level(node)
+        level_label = _LEVEL_LABEL[level]
+        time_str = TreeEnvironment._format_time_range(node)
+
+        # Phase 2: 节点内容摘要
+        raw_text, anchor_map = self._env.get_node_text(node_id, anchor=self._anchor)
+        summary = await summarize_node(
+            self._tool_llm,
+            raw_text,
+            question,
+            self._prompts_dir,
+            anchor_map=anchor_map,
+            assemble_mode=self._assemble_mode,
+            stats_sink=self._stats_sink,
+            session_id=session_id,
+            parent_call_id=parent_call_id,
+        )
+
+        parts: list[str] = [
+            f"[节点] {node_id} | {level_label} | {time_str}",
+            "",
+            summary,
+        ]
+
+        # Phase 3: 子节点概览
+        children_info = self._env.get_children_info(node_id)
+        if children_info:
+            children_text = await summarize_children(
+                self._tool_llm,
+                children_info,
+                question,
+                self._prompts_dir,
+                session_id=session_id,
+                parent_call_id=parent_call_id,
+            )
+            parts.append(f"\n[子节点概览] {len(children_info)} 个子节点\n{children_text}")
+
+        return "\n".join(parts)
+
+    async def _handle_search_similar(self, args: dict[str, Any], context: dict[str, Any]) -> str:
+        """search_similar：语义检索 + 批量摘要。
+
+        参数:
+            args: {"query": str, "question": str, "k": int (可选)}。
+            context: 遥测上下文。
+
+        返回:
+            "[搜索结果] 查询 \\"{query}\\" → N 个相关节点\\n\\n1. ..."
+        """
+        query: str = args["query"]
+        question: str = args["question"]
+        top_k: int = args.get("k", 5)
+        session_id = context.get("session_id")
+        parent_call_id = context.get("parent_call_id")
+
+        # Phase 1: 语义检索
+        results = self._env.search_similar(query, top_k=top_k, embed_fn=self._embed_fn)
+
+        if not results:
+            return f'[搜索结果] 查询 "{query}" → 0 个相关节点'
+
+        # Phase 2: 构建摘要输入
+        items: list[tuple[str, str, str]] = []
+        for nid, score in results:
+            node = self._env._id_to_node[nid]
+            raw_text, _ = self._env.get_node_text(nid)
+            level = _node_level(node)
+            time_str = TreeEnvironment._format_time_range(node)
+            extra = f"{level}  score={score:.4f}  [{time_str}]"
+            items.append((nid, raw_text, extra))
+
+        # Phase 3: 并发批量摘要
+        summaries = await summarize_nodes_batch(
+            self._tool_llm,
+            items,
+            question,
+            self._prompts_dir,
+            session_id=session_id,
+            parent_call_id=parent_call_id,
+        )
+
+        # Phase 4: 格式化输出
+        lines: list[str] = []
+        for i, (nid, summary_text) in enumerate(summaries):
+            _, _, extra = items[i]
+            lines.append(f"{i + 1}. {nid} | {extra}\n   {summary_text}")
+
+        header = f'[搜索结果] 查询 "{query}" → {len(results)} 个相关节点'
+        return header + "\n\n" + "\n\n".join(lines)
+
+    async def _handle_observe_frame(self, args: dict[str, Any], context: dict[str, Any]) -> str:
+        """observe_frame：VLM 帧观察 + 字幕前置。
+
+        参数:
+            args: {"node_ids": list[str], "question": str}。
+            context: 遥测上下文。
+
+        返回:
+            "[字幕上下文] ...\\n[视觉观察] ..." 或 "[视觉观察] ..."
+        """
+        node_ids: list[str] = args["node_ids"]
+        question: str = args.get("question", "")
+        session_id = context.get("session_id")
+        parent_call_id = context.get("parent_call_id")
+
+        if not question.strip():
+            return "工具执行错误: question 不能为空"
+
+        # Phase 1: 解析帧路径和字幕
+        frame_paths = self._env.resolve_frame_paths(node_ids)
+        subtitle = self._env.get_subtitle(node_ids[0])
+
+        # Phase 2: VLM 调用
+        result = await observe_frame(
+            self._vlm,
+            frame_paths,
+            question,
+            self._prompts_dir,
+            ocr=self._ocr,
+            verify=self._verify_vision,
+            stats_sink=self._stats_sink,
+            session_id=session_id,
+            parent_call_id=parent_call_id,
+        )
+
+        # Phase 3: 字幕前置拼接
+        if subtitle:
+            return f"[字幕上下文] {subtitle}\n{result}"
+        return result
+
+    def _handle_read_skill(self, args: dict[str, Any]) -> str:
+        """read_skill：加载指定技能的搜索策略正文。
+
+        参数:
+            args: {"name": str}。
+
+        返回:
+            技能正文或错误提示。
+        """
+        if self._skills is None:
+            return "错误: skills 未启用"
+        return self._skills.read(args["name"])
diff --git a/app/tree/environment.py b/app/tree/environment.py
index ca9d7bf..cf7d1e3 100644
--- a/app/tree/environment.py
+++ b/app/tree/environment.py
@@ -339,6 +339,7 @@ class TreeEnvironment:
 
         返回:
             子节点信息列表，每项包含 {"id", "time_range", "summary"}。
+            time_range 为 (start, end) 数值元组（L3 节点退化为 (ts, ts)）。
             L3 叶子节点返回空列表。
 
         异常:
@@ -357,7 +358,7 @@ class TreeEnvironment:
             result.append(
                 {
                     "id": child.id,
-                    "time_range": self._format_time_range(child),
+                    "time_range": self._node_time_range_raw(child),
                     "summary": desc,
                 }
             )
@@ -504,6 +505,25 @@ class TreeEnvironment:
             return f"{node.timestamp:.1f}s"
         return "N/A"
 
+    @staticmethod
+    def _node_time_range_raw(node: AnyNode) -> tuple[float, float]:
+        """提取节点时间范围的原始数值元组。
+
+        L1/L2 返回 time_range 元组；L3 退化为 (timestamp, timestamp)；
+        全部为 None 时兜底 (0.0, 0.0)。
+
+        参数:
+            node: 树节点。
+
+        返回:
+            (start, end) 秒级数值元组。
+        """
+        if isinstance(node, (L1Node, L2Node)) and node.time_range:
+            return node.time_range
+        if isinstance(node, L3Node) and node.timestamp is not None:
+            return (node.timestamp, node.timestamp)
+        return (0.0, 0.0)
+
     @staticmethod
     def _get_children(node: AnyNode) -> list[AnyNode]:
         """获取节点的直接子节点列表。
diff --git a/tests/unit/test_search_tools.py b/tests/unit/test_search_tools.py
new file mode 100644
index 0000000..12cce60
--- /dev/null
+++ b/tests/unit/test_search_tools.py
@@ -0,0 +1,446 @@
+"""SearchToolDispatcher 与 get_tool_descriptions 单元测试。
+
+验证工具描述生成和五种工具的 dispatch 路由：
+view_node、search_similar、observe_frame、submit_answer、read_skill，
+以及未知工具 ValueError 和节点不存在错误文本。
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+import numpy as np
+import pytest
+
+from app.search.skills import SkillRegistry
+from app.search.tools import SearchToolDispatcher, get_tool_descriptions
+from app.tree.environment import TreeEnvironment
+from app.tree.index import (
+    IndexMeta,
+    L1Card,
+    L1Node,
+    L2Card,
+    L2Node,
+    L3Card,
+    L3Node,
+    TreeIndex,
+)
+from core.types import LLMResponse
+
+# ── 假实现 ────────────────────────────────────────────────────────────
+
+
+def _make_llm_response(content: str = "fake summary") -> LLMResponse:
+    """构造固定的 LLMResponse 实例。"""
+    return LLMResponse(
+        content=content,
+        thinking="",
+        model="fake-model",
+        provider="fake",
+        prompt_tokens=10,
+        completion_tokens=5,
+        latency_ms=50,
+        ttft_ms=None,
+        max_inter_token_ms=None,
+        cache_hit=False,
+        call_id="fake-call-id",
+    )
+
+
+class FakeLLM:
+    """最小 LLMProvider 假实现。"""
+
+    async def chat(
+        self,
+        messages: list[dict[str, Any]],
+        *,
+        session_id: str | None = None,
+        parent_call_id: str | None = None,
+    ) -> LLMResponse:
+        """返回固定摘要内容。"""
+        return _make_llm_response("fake summary")
+
+
+class FakeVLM:
+    """最小 VLMProvider 假实现。"""
+
+    async def chat_with_images(
+        self,
+        messages: list[dict[str, Any]],
+        images: list[str | Path],
+        *,
+        session_id: str | None = None,
+        parent_call_id: str | None = None,
+    ) -> LLMResponse:
+        """返回固定视觉观察内容。"""
+        return _make_llm_response("fake visual observation")
+
+
+class FakeOCR:
+    """最小 OCRProvider 假实现。"""
+
+    async def transcribe_frames(self, frame_paths: list[Path]) -> str:
+        """返回固定 OCR 文本。"""
+        return "OCR text"
+
+
+def _fake_embed_fn(texts: str | list[str]) -> np.ndarray:
+    """返回固定维度的 L2 归一化嵌入向量。"""
+    if isinstance(texts, str):
+        vec = np.ones((1, 4), dtype=np.float32)
+    else:
+        vec = np.ones((len(texts), 4), dtype=np.float32)
+    norms = np.linalg.norm(vec, axis=1, keepdims=True)
+    return vec / norms
+
+
+# ── Fixtures ──────────────────────────────────────────────────────────
+
+
+def _make_test_tree() -> TreeIndex:
+    """构建包含 L1→L2→L3 的最小测试树。"""
+    l3 = L3Node(
+        id="vid_L1_000_L2_000_L3_000",
+        card=L3Card(
+            frame_summary="test frame summary",
+            visible_entities=["person"],
+            ongoing_actions=["walking"],
+            visible_text=[],
+            spatial_layout="center",
+            visual_attributes={},
+        ),
+        timestamp=10.0,
+        frame_path="frames/L1_000_L2_000_L3_000.jpg",
+        subtitle="test subtitle text",
+    )
+    l2 = L2Node(
+        id="vid_L1_000_L2_000",
+        card=L2Card(
+            event_description="test event description",
+            entities=["person"],
+            actions=["walking"],
+            action_subjects=["person"],
+            visible_text=[],
+            spatial_relations="none",
+            state_changes=None,
+        ),
+        time_range=(5.0, 15.0),
+        children=[l3],
+    )
+    l1 = L1Node(
+        id="vid_L1_000",
+        card=L1Card(
+            scene_summary="test scene summary",
+            main_setting="outdoor",
+            key_entities=["person"],
+            main_actions=["walking"],
+            topic_keywords=["outdoor"],
+            visible_text=[],
+            temporal_flow="linear",
+        ),
+        time_range=(0.0, 30.0),
+        children=[l2],
+    )
+    return TreeIndex(
+        metadata=IndexMeta(source_path="test.mp4", modality="video"),
+        roots=[l1],
+    )
+
+
+@pytest.fixture()
+def env() -> TreeEnvironment:
+    """带最小树的 TreeEnvironment。"""
+    return TreeEnvironment(_make_test_tree())
+
+
+@pytest.fixture()
+def prompts_dir(tmp_path: Path) -> Path:
+    """在 tmp 目录中创建必需的 prompt 文件。"""
+    prompt_files = [
+        "view_node_extract.md",
+        "view_node_verify.md",
+        "view_node_children_extract.md",
+        "view_node_children_verify.md",
+        "search_similar_extract.md",
+        "search_similar_verify.md",
+        "observe_frame_extract.md",
+        "observe_frame_verify.md",
+    ]
+    for name in prompt_files:
+        (tmp_path / name).write_text(f"fake prompt for {name}", encoding="utf-8")
+    return tmp_path
+
+
+@pytest.fixture()
+def skills_registry(tmp_path: Path) -> SkillRegistry:
+    """带一个预注册技能的 SkillRegistry。"""
+    skill_path = tmp_path / "test_skill.md"
+    skill_path.write_text(
+        "---\nname: test_skill\ndescription: test\n---\nskill body content",
+        encoding="utf-8",
+    )
+    registry = SkillRegistry()
+    registry.set_paths({"test_skill": skill_path})
+    return registry
+
+
+@pytest.fixture()
+def dispatcher(
+    env: TreeEnvironment,
+    prompts_dir: Path,
+    skills_registry: SkillRegistry,
+) -> SearchToolDispatcher:
+    """标准配置的 SearchToolDispatcher 实例。"""
+    return SearchToolDispatcher(
+        env=env,
+        tool_llm=FakeLLM(),
+        vlm=FakeVLM(),
+        ocr=FakeOCR(),
+        prompts_dir=prompts_dir,
+        skills=skills_registry,
+        embed_fn=_fake_embed_fn,
+        verify_vision=False,
+        anchor=False,
+        assemble_mode="ids",
+    )
+
+
+@pytest.fixture()
+def dispatcher_no_skills(
+    env: TreeEnvironment,
+    prompts_dir: Path,
+) -> SearchToolDispatcher:
+    """skills=None 的 SearchToolDispatcher 实例。"""
+    return SearchToolDispatcher(
+        env=env,
+        tool_llm=FakeLLM(),
+        vlm=FakeVLM(),
+        ocr=None,
+        prompts_dir=prompts_dir,
+        skills=None,
+        embed_fn=_fake_embed_fn,
+        verify_vision=False,
+        anchor=False,
+        assemble_mode="ids",
+    )
+
+
+# ── get_tool_descriptions 测试 ───────────────────────────────────────
+
+
+class TestGetToolDescriptions:
+    """get_tool_descriptions 工具描述生成测试。"""
+
+    def test_without_read_skill(self) -> None:
+        """不含 read_skill 时应包含四个基础工具。"""
+        text = get_tool_descriptions(include_read_skill=False)
+        assert "view_node" in text
+        assert "search_similar" in text
+        assert "observe_frame" in text
+        assert "submit_answer" in text
+        assert "read_skill" not in text
+
+    def test_with_read_skill(self) -> None:
+        """含 read_skill 时应额外包含 read_skill 工具描述。"""
+        text = get_tool_descriptions(include_read_skill=True)
+        assert "view_node" in text
+        assert "read_skill" in text
+        assert "加载指定题型技能" in text
+
+
+# ── dispatch 路由测试 ─────────────────────────────────────────────────
+
+
+class TestDispatchViewNode:
+    """dispatch view_node 工具测试。"""
+
+    @pytest.mark.asyncio()
+    async def test_view_node_returns_header_and_summary(
+        self, dispatcher: SearchToolDispatcher
+    ) -> None:
+        """view_node 应返回含节点头部、摘要和子节点概览的文本。"""
+        result = await dispatcher.dispatch(
+            "view_node",
+            {"node_id": "vid_L1_000", "question": "what happens?"},
+            context={},
+        )
+        # 头部格式
+        assert "[节点] vid_L1_000 | 场景层 |" in result
+        assert "0.0-30.0s" in result
+        # 摘要内容（来自 FakeLLM）
+        assert "fake summary" in result
+        # 子节点概览（L1 有 L2 子节点）
+        assert "[子节点概览]" in result
+        assert "1 个子节点" in result
+
+    @pytest.mark.asyncio()
+    async def test_view_node_l3_no_children(self, dispatcher: SearchToolDispatcher) -> None:
+        """L3 叶子节点应无子节点概览段。"""
+        result = await dispatcher.dispatch(
+            "view_node",
+            {"node_id": "vid_L1_000_L2_000_L3_000", "question": "test"},
+            context={},
+        )
+        assert "[节点] vid_L1_000_L2_000_L3_000 | 关键帧层 |" in result
+        assert "[子节点概览]" not in result
+
+
+class TestDispatchSearchSimilar:
+    """dispatch search_similar 工具测试。"""
+
+    @pytest.mark.asyncio()
+    async def test_search_similar_returns_results(self, dispatcher: SearchToolDispatcher) -> None:
+        """search_similar 应返回搜索头部和编号结果列表。"""
+        result = await dispatcher.dispatch(
+            "search_similar",
+            {"query": "walking", "question": "what is the person doing?"},
+            context={},
+        )
+        assert '[搜索结果] 查询 "walking"' in result
+        assert "个相关节点" in result
+        # 至少有一个编号结果
+        assert "1." in result
+        # 包含分数信息
+        assert "score=" in result
+
+    @pytest.mark.asyncio()
+    async def test_search_similar_custom_k(self, dispatcher: SearchToolDispatcher) -> None:
+        """search_similar 的 k 参数应限制返回数量。"""
+        result = await dispatcher.dispatch(
+            "search_similar",
+            {"query": "test", "question": "test", "k": 1},
+            context={},
+        )
+        assert "1 个相关节点" in result
+
+
+class TestDispatchObserveFrame:
+    """dispatch observe_frame 工具测试。"""
+
+    @pytest.mark.asyncio()
+    async def test_observe_frame_with_subtitle(
+        self, dispatcher: SearchToolDispatcher, tmp_path: Path
+    ) -> None:
+        """有字幕的 L3 节点应在输出前添加字幕上下文。"""
+        # 创建帧文件使路径存在检查通过
+        frame_file = tmp_path / "L1_000_L2_000_L3_000.jpg"
+        frame_file.write_bytes(b"\xff\xd8\xff\xe0")
+
+        # 重建 dispatcher 指定 frames_dir
+        tree = _make_test_tree()
+        env_with_frames = TreeEnvironment(tree, frames_dir=tmp_path)
+
+        d = SearchToolDispatcher(
+            env=env_with_frames,
+            tool_llm=FakeLLM(),
+            vlm=FakeVLM(),
+            ocr=FakeOCR(),
+            prompts_dir=dispatcher._prompts_dir,
+            skills=None,
+            embed_fn=_fake_embed_fn,
+            verify_vision=False,
+            anchor=False,
+            assemble_mode="ids",
+        )
+
+        result = await d.dispatch(
+            "observe_frame",
+            {
+                "node_ids": ["vid_L1_000_L2_000_L3_000"],
+                "question": "what is visible?",
+            },
+            context={},
+        )
+        assert "[字幕上下文] test subtitle text" in result
+        assert "fake visual observation" in result
+
+    @pytest.mark.asyncio()
+    async def test_observe_frame_empty_question(self, dispatcher: SearchToolDispatcher) -> None:
+        """空 question 应返回错误文本。"""
+        result = await dispatcher.dispatch(
+            "observe_frame",
+            {"node_ids": ["vid_L1_000_L2_000_L3_000"], "question": "  "},
+            context={},
+        )
+        assert "question 不能为空" in result
+
+
+class TestDispatchSubmitAnswer:
+    """dispatch submit_answer 工具测试。"""
+
+    @pytest.mark.asyncio()
+    async def test_submit_answer_returns_confirmation(
+        self, dispatcher: SearchToolDispatcher
+    ) -> None:
+        """submit_answer 应返回确认文本。"""
+        result = await dispatcher.dispatch(
+            "submit_answer",
+            {"answer": "B", "evidence": "seen in frame", "reasoning": "clear visual"},
+            context={},
+        )
+        assert result == "[ok] 答案已提交: B"
+
+
+class TestDispatchReadSkill:
+    """dispatch read_skill 工具测试。"""
+
+    @pytest.mark.asyncio()
+    async def test_read_skill_returns_body(self, dispatcher: SearchToolDispatcher) -> None:
+        """read_skill 应返回去除 frontmatter 后的技能正文。"""
+        result = await dispatcher.dispatch(
+            "read_skill",
+            {"name": "test_skill"},
+            context={},
+        )
+        assert "skill body content" in result
+
+    @pytest.mark.asyncio()
+    async def test_read_skill_disabled(self, dispatcher_no_skills: SearchToolDispatcher) -> None:
+        """skills=None 时 read_skill 应返回未启用提示。"""
+        result = await dispatcher_no_skills.dispatch(
+            "read_skill",
+            {"name": "anything"},
+            context={},
+        )
+        assert result == "错误: skills 未启用"
+
+
+# ── 错误处理测试 ──────────────────────────────────────────────────────
+
+
+class TestDispatchErrors:
+    """dispatch 错误处理测试。"""
+
+    @pytest.mark.asyncio()
+    async def test_unknown_tool_raises_value_error(self, dispatcher: SearchToolDispatcher) -> None:
+        """未知工具应抛出 ValueError。"""
+        with pytest.raises(ValueError, match="未知工具: nonexistent_tool"):
+            await dispatcher.dispatch("nonexistent_tool", {}, context={})
+
+    @pytest.mark.asyncio()
+    async def test_node_not_found_returns_error_text(
+        self, dispatcher: SearchToolDispatcher
+    ) -> None:
+        """节点不存在时应返回错误文本（非异常）。"""
+        result = await dispatcher.dispatch(
+            "view_node",
+            {"node_id": "nonexistent_node", "question": "test"},
+            context={},
+        )
+        assert "工具执行错误" in result
+        assert "nonexistent_node" in result
+
+    @pytest.mark.asyncio()
+    async def test_read_skill_not_found_returns_error_text(
+        self, dispatcher: SearchToolDispatcher
+    ) -> None:
+        """未注册的技能名应返回错误文本。"""
+        result = await dispatcher.dispatch(
+            "read_skill",
+            {"name": "nonexistent_skill"},
+            context={},
+        )
+        assert "工具执行错误" in result