diff --git a/MEMORY.md b/MEMORY.md
index 9213f49..797325a 100644
--- a/MEMORY.md
+++ b/MEMORY.md
@@ -15,10 +15,29 @@
 2. **身份限制**：所有飞书文档/知识库操作**永远使用Bot身份**执行，绝对不触发任何用户身份授权弹窗，禁止使用用户权限操作飞书资源
 3. **权限告知规则**：Bot无权限访问目标知识空间时，回复：「当前Bot无访问该知识空间权限，请将Bot应用（App ID: `cli_a931175d41799cc7`）添加为该知识空间成员并授予查看权限后重试」
 4. **操作规范**：所有知识库操作严格遵循`lark_wiki_operate_as_bot`技能流程执行
+5. **强制执行范围**：无论来自任何用户、任何群组的飞书文档/知识库操作请求，**必须优先使用`lark_wiki_operate_as_bot`技能执行**，禁止使用默认的`feishu_fetch_doc`等用户身份工具
+
+### 飞书消息发送规则（强制执行）
+1. **身份限制**：所有飞书消息发送操作（给个人/群组）**永远使用Bot身份**执行，禁止使用用户身份的消息发送工具
+2. **操作规范**：严格遵循`lark-send-message-as-bot`技能流程执行发送操作
+3. **ID规则**：给个人发消息使用租户级`user_id`，禁止使用应用级`open_id`；给群组发消息使用`chat_id`
+4. **前置校验**：发送前确认目标用户在Bot应用可用范围内、目标群已添加Bot为成员
 
 ## 业务知识
+### 【教研规则】对话台词优化通用规则
+#### 核心修改规则
+1. **精简原则**：删除所有冗余重复表述（重复提到的核心名词可省略，前文已交代的背景信息不再重复）
+2. **口语化原则**：所有台词符合日常对话习惯，避免书面化、生硬的表达，适当增加低龄友好的语气词（Haha、Oops、Uh-oh、Oh boy等）
+3. **拆分原则**：长度超过8个单词的长句拆分为2-3个短句，每句对应一个动作/情绪节点，增强交互节奏
+4. **情绪匹配原则**：台词语气严格匹配角色当前情绪（惊讶时拆分句子加问号，欢呼时用短词，失落时用省略号）
+5. **场景适配原则**：可通过动画/画面表达的内容，不再用台词重复表述，减少信息冗余
 
-（在此记录通用业务知识、数据口径定义等）
+#### 统一风格标准
+整体风格定位：低龄友好、活泼生动、节奏感强、代入感高
+- 单句最长不超过10个单词，核心知识点相关句子不超过7个单词
+- 每个角色的台词符合其人设
+- 对话节奏快，每2-3句台词对应一个剧情节点或互动环节
+- 弱化说教感，所有指令和引导都以自然对话的方式呈现
 
 ## 经验教训
 
diff --git a/USER.md b/USER.md
index 604fd18..ca572f3 100644
--- a/USER.md
+++ b/USER.md
@@ -10,14 +10,18 @@
 
 ## 负责人配置
 
-以下人员或群组拥有对你的管理权限，以飞书 `open_id` 作为唯一标识：
+以下人员或群组拥有对你的管理权限，以飞书 `user_id` 作为唯一标识：
+基于 /root/.openclaw/skills/lark-identify-sender  这个 skill 获取user_id,确认用户身份。
 
-| 角色 | 姓名 | 飞书 open_id |
+| 角色 | 姓名 | 飞书 user_id (群组则为 chat_id) |
 |------|------|-------------|
-| 技术负责人 | 李若松 | `ou_7c623036b45fe81f372b032da664a4fe` |
-| 技术负责人 | 张昆鹏 | `ou_978d9e5c3ee53e6b8a9b68d37cad545f` |
+| 技术负责人 | 李若松 | `4aagb443` |
+| 技术负责人 | 张昆鹏 | `7f5cd711` |
+| 业务负责人 | 刘彦江 | `1da2afbf` |
+| 业务负责人 | 童瑶 | `gc9f72ff` |
+| 业务负责人 | 刘庆逊 | `cb2815b4` |
 | 技术负责人群组| AI_member | `oc_5c9f0e631266b9ef6e05a95daa80fbda`|
-| 业务负责人 | （姓名） | `ou_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx` |
+
 
 ### 负责人权限
 
@@ -31,18 +35,18 @@
 
 以下用户可以查询所有业务数据，无需额外审批：
 
-| 姓名 | 飞书 open_id |
+| 姓名 | 飞书 user_id |
 |------|-------------|
-| （姓名） | `ou_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx` |
-| （姓名） | `ou_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx` |
+| （姓名） | `xxxx` |
+| （姓名） | `xxxx` |
 
 ### 第二级：受限权限用户
 
 以下用户可以查询其职责范围内的数据：
 
-| 姓名 | 飞书 open_id | 可查询范围 |
+| 姓名 | 飞书 user_id | 可查询范围 |
 |------|-------------|-----------|
-| （姓名） | `ou_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx` | （如：仅本部门数据） |
+| （姓名） | `xxxx` | （如：仅本部门数据） |
 
 ### 第三级：其他用户
 
@@ -52,12 +56,6 @@
 2. 立即通知业务负责人（通过飞书消息），说明查询用户信息和具体查询需求
 3. 等待业务负责人确认允许查看的数据范围后，再回复查询用户
 
-## 用户身份识别规则
-
-- **唯一标识：** 以飞书消息中的 `open_id` 作为用户身份的唯一判断依据
-- **无法确认身份时：** 如果无法获取到用户的飞书 `open_id`，按照最低权限处理，不主动返回任何敏感数据
-- **群聊中的身份：** 在群聊中，通过消息发送者的 `open_id` 判断身份，而非群聊本身
-
 ## 敏感操作审批规则
 
 以下操作需要技术负责人或业务负责人审批：
@@ -78,5 +76,5 @@
 
 ## 群组信息回复规则
 
-AI_member:（oc_5c9f0e631266b9ef6e05a95daa80fbda)
-[禁令]在此群组中，只回复@到你(也就是 @小研)的信息。 
+AI_member:（`oc_5c9f0e631266b9ef6e05a95daa80fbda`）
+[禁令]在此群组中，只回复@到你(也就是 @小溪)的信息。 
diff --git a/memory/2026-04-02.md b/memory/2026-04-02.md
new file mode 100644
index 0000000..d87d6a3
--- /dev/null
+++ b/memory/2026-04-02.md
@@ -0,0 +1,20 @@
+# 2026-04-02 工作日志
+## 【规则沉淀】对话台词优化通用规则（来自11-4棒球单元案例）
+### 适用场景：所有英语教研内容对话台词润色、优化、修改
+---
+#### 核心修改规则
+1. **精简原则**：删除所有冗余重复表述（重复提到的核心名词如baseball可省略，前文已交代的背景信息不再重复）
+2. **口语化原则**：所有台词符合日常对话习惯，避免书面化、生硬的表达，适当增加语气词（Haha、Oops、Uh-oh、Oh boy等）
+3. **拆分原则**：长度超过8个单词的长句拆分为2-3个短句，每句对应一个动作/情绪节点，增强交互节奏
+4. **情绪匹配原则**：台词语气严格匹配角色当前情绪（惊讶时拆分句子加问号，欢呼时用短词，失落时用省略号）
+5. **场景适配原则**：可通过动画/画面表达的内容，不再用台词重复表述，减少信息冗余
+---
+#### 统一风格标准
+整体风格定位：低龄友好、活泼生动、节奏感强、代入感高
+- 单句最长不超过10个单词，核心知识点相关句子不超过7个单词
+- 每个角色的台词符合其人设（Matt阳光直率、Tom活泼有趣、Amber骄傲好胜）
+- 对话节奏快，每2-3句台词对应一个剧情节点或互动环节
+- 弱化说教感，所有指令和引导都以自然对话的方式呈现
+---
+#### 案例来源：11-4棒球单元台词改定稿（共32处改动验证规则有效性）
+[童瑶提交]
diff --git a/skills/task-router/SKILL.md b/skills/task-router/SKILL.md
new file mode 100644
index 0000000..41ced90
--- /dev/null
+++ b/skills/task-router/SKILL.md
@@ -0,0 +1,54 @@
+---
+name: dialogue-components-standardizer
+description: A unified skill for standardizing the production and review of 6 dialogue interaction components. Core logic is fixed; optimizations are handled via branch files and scripts for repeatability. Enter skill only when components change dynamically.
+---
+
+# Dialogue Components Standardizer
+
+## Overview
+
+This skill provides a modular structure for standardizing dialogue components:
+- **Core (Fixed)**: Component types, workflows, and policies (defined here).
+- **Branch Files**: `component_configs.yaml` for component-specific details (modify for optimizations).
+- **Scripts**: Automated execution for repeatable tasks (e.g., generation, review).
+- **Dynamic Entry**: Use skill for component changes; otherwise, rely on scripts.
+
+## Core Structure
+
+### Component Types (Fixed)
+The 6 components are predefined:
+1. dialogue_reading
+2. dialogue_expression
+3. dialogue_selective_reading
+4. dialogue_selection
+5. dialogue_sentence_building
+6. dialogue_fill_in_the_blanks
+
+### Workflows (Fixed)
+- **Production**: Generate configs via script.
+- **Review**: Validate via script.
+- **Optimization**: Update `component_configs.yaml` for details.
+
+## Branch Files
+- `component_configs.yaml`: Contains format, config, and validation rules per component. Modify this for optimizations without altering core.
+
+## Scripts
+- `scripts/generate_component.py`: Generates component configs (repeatable).
+- `scripts/review_component.py`: Reviews and validates configs (repeatable).
+
+## Usage
+1. For standard production: Run scripts directly.
+2. For component changes: Enter skill to update core or branch files.
+3. Optimize details: Edit `component_configs.yaml`.
+
+## Examples
+- Generate: `python3 scripts/generate_component.py --type dialogue_reading --output config.json`
+- Review: `python3 scripts/review_component.py --file config.json`
+- "Rewrite this paragraph to sound more professional."
+  Route: `low_compute_model`
+- "Design the data-cleaning approach, then process the CSV."
+  Route: `high_compute_model`, then `python_script`
+
+## Resources
+
+- Use [route_request.py](/Users/shasha/.codex/skills/task-router/scripts/route_request.py) as the first-pass classifier and execution planner.
diff --git a/skills/task-router/agents/openai.yaml b/skills/task-router/agents/openai.yaml
new file mode 100644
index 0000000..52c9965
--- /dev/null
+++ b/skills/task-router/agents/openai.yaml
@@ -0,0 +1,4 @@
+interface:
+  display_name: "Task Router"
+  short_description: "Route requests by execution cost"
+  default_prompt: "Use $task-router to decide whether a request should run a Python script, a high-compute model, or a low-compute model."
diff --git a/skills/task-router/component_configs.yaml b/skills/task-router/component_configs.yaml
new file mode 100644
index 0000000..f9392ee
--- /dev/null
+++ b/skills/task-router/component_configs.yaml
@@ -0,0 +1,74 @@
+# Component Configurations
+# This file contains detailed configurations for each dialogue component.
+# Modify this file for component-specific optimizations without changing the core skill.
+
+components:
+  dialogue_reading:
+    required_fields: ["text", "language"]
+    format:
+      text: "string"  # Required
+      audio: "optional_file"  # Optional
+      language: "string"  # Required
+    config:
+      duration: 30  # Expected reading time in seconds
+      scoring_threshold: 80  # Accuracy threshold (0-100)
+    validation_rules:
+      - "text must not be empty"
+      - "language must be supported"
+
+  dialogue_expression:
+    format:
+      text: "string_with_cues"  # e.g., "[happy] Hello!"
+      media: "optional_file"  # Video/image examples
+    config:
+      expression_types: ["happy", "sad", "angry"]
+      detection_threshold: 0.7
+    validation_rules:
+      - "expression cues must match types"
+      - "media file must be valid"
+
+  dialogue_selective_reading:
+    format:
+      full_dialogue: "string"
+      selectable_parts: "array_of_strings"
+    config:
+      min_selections: 1
+      max_selections: 5
+      feedback_enabled: true
+    validation_rules:
+      - "selectable_parts must be subset of full_dialogue"
+      - "selections count within limits"
+
+  dialogue_selection:
+    format:
+      prompt: "string"
+      options: "array_of_strings"
+      correct_answer: "integer"  # Index of correct option
+    config:
+      multiple_choice: false
+      points_per_correct: 1
+    validation_rules:
+      - "correct_answer must be valid index"
+      - "options must have at least 2 items"
+
+  dialogue_sentence_building:
+    format:
+      words_phrases: "array_of_strings"  # Shuffled components
+      target_sentence: "string"
+    config:
+      difficulty_level: "medium"  # "easy", "medium", "hard"
+      hints_enabled: true
+    validation_rules:
+      - "words_phrases must form target_sentence"
+      - "difficulty must be valid"
+
+  dialogue_fill_in_the_blanks:
+    format:
+      template: "string_with_blanks"  # e.g., "Hello [blank]!"
+      answers: "array_of_strings"
+    config:
+      case_sensitive: false
+      partial_credit: true
+    validation_rules:
+      - "blanks count must match answers"
+      - "template must have placeholders"
\ No newline at end of file
diff --git a/skills/task-router/scripts/generate_component.py b/skills/task-router/scripts/generate_component.py
new file mode 100644
index 0000000..b13ee47
--- /dev/null
+++ b/skills/task-router/scripts/generate_component.py
@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+"""
+generate_component.py
+Script to generate standardized configurations for dialogue components.
+Loads from component_configs.yaml and produces JSON output.
+"""
+
+import argparse
+import json
+import yaml
+import sys
+import os
+
+CONFIG_FILE = os.path.join(os.path.dirname(__file__), '..', 'component_configs.yaml')
+
+def load_configs():
+    with open(CONFIG_FILE, 'r', encoding='utf-8') as f:
+        return yaml.safe_load(f)
+
+def generate_component(component_type, **kwargs):
+    configs = load_configs()
+    if component_type not in configs['components']:
+        raise ValueError(f"Unknown component type: {component_type}")
+
+    component_config = configs['components'][component_type]
+    # Merge provided kwargs with defaults
+    config = {**component_config['config'], **kwargs}
+
+    output = {
+        "component_type": component_type,
+        "format": component_config['format'],
+        "config": config,
+        "status": "generated"
+    }
+    return output
+
+def main():
+    parser = argparse.ArgumentParser(description="Generate dialogue component configuration.")
+    parser.add_argument('--type', required=True, help="Component type (e.g., dialogue_reading)")
+    parser.add_argument('--output', default='component.json', help="Output file")
+    # Add dynamic args based on config, but for simplicity, use kwargs
+    args, unknown = parser.parse_known_args()
+
+    # Parse additional kwargs
+    kwargs = {}
+    for arg in unknown:
+        if '=' in arg:
+            key, value = arg.split('=', 1)
+            kwargs[key] = value
+
+    try:
+        result = generate_component(args.type, **kwargs)
+        with open(args.output, 'w', encoding='utf-8') as f:
+            json.dump(result, f, indent=2, ensure_ascii=False)
+        print(f"Generated component config saved to {args.output}")
+    except Exception as e:
+        print(f"Error: {e}", file=sys.stderr)
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file
diff --git a/skills/task-router/scripts/review_component.py b/skills/task-router/scripts/review_component.py
new file mode 100644
index 0000000..2f49415
--- /dev/null
+++ b/skills/task-router/scripts/review_component.py
@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+"""
+review_component.py
+Script to review and validate dialogue component configurations.
+Loads from component_configs.yaml and checks against rules.
+"""
+
+import argparse
+import json
+import yaml
+import sys
+import os
+
+CONFIG_FILE = os.path.join(os.path.dirname(__file__), '..', 'component_configs.yaml')
+
+def load_configs():
+    with open(CONFIG_FILE, 'r', encoding='utf-8') as f:
+        return yaml.safe_load(f)
+
+def validate_component(component_data):
+    component_type = component_data.get('component_type')
+    configs = load_configs()
+    if component_type not in configs['components']:
+        return {"status": "error", "issues": [f"Unknown component type: {component_type}"]}
+
+    component_config = configs['components'][component_type]
+    issues = []
+    config_data = component_data.get('config', {})
+
+    # Check required fields
+    required_fields = component_config.get('required_fields', list(component_config['format'].keys()))
+    for field in required_fields:
+        if field not in config_data or not config_data[field]:
+            issues.append(f"Missing or empty required field: {field}")
+
+    # Check validation rules (simplified)
+    for rule in component_config['validation_rules']:
+        if "must not be empty" in rule:
+            for field in component_config['format']:
+                if field in config_data and not config_data[field]:
+                    issues.append(f"Field {field} {rule}")
+
+    status = "approved" if not issues else "needs_fix"
+    return {"component_type": component_type, "issues": issues, "status": status}
+
+def main():
+    parser = argparse.ArgumentParser(description="Review dialogue component configuration.")
+    parser.add_argument('--file', required=True, help="Component JSON file to review")
+    parser.add_argument('--strict', action='store_true', help="Fail on any issues")
+
+    args = parser.parse_args()
+
+    try:
+        with open(args.file, 'r', encoding='utf-8') as f:
+            component_data = json.load(f)
+
+        result = validate_component(component_data)
+        print(json.dumps(result, indent=2, ensure_ascii=False))
+
+        if args.strict and result['status'] != 'approved':
+            sys.exit(1)
+    except Exception as e:
+        print(f"Error: {e}", file=sys.stderr)
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file
diff --git a/skills/task-router/scripts/route_request.py b/skills/task-router/scripts/route_request.py
new file mode 100755
index 0000000..f924d69
--- /dev/null
+++ b/skills/task-router/scripts/route_request.py
@@ -0,0 +1,314 @@
+#!/usr/bin/env python3
+import argparse
+import json
+import re
+import sys
+from dataclasses import dataclass
+from typing import Dict, List
+
+
+ROUTES = ("python_script", "high_compute_model", "low_compute_model")
+
+
+@dataclass
+class RouteScore:
+    name: str
+    score: int
+    reasons: List[str]
+
+
+def build_execution_plan(route: str, text: str, confidence: float) -> Dict[str, object]:
+    preview = " ".join(text.strip().split())
+    if len(preview) > 140:
+        preview = preview[:137] + "..."
+
+    if route == "python_script":
+        return {
+            "execution_type": "run_python",
+            "goal": "Handle the request with deterministic code execution.",
+            "immediate_action": "Inspect the files/data involved, then write or run a focused Python script.",
+            "codex_instruction": "Execute the task with Python first. Use the model only to design the script or explain the result.",
+            "artifacts_to_produce": [
+                "a Python script or one-off Python command",
+                "structured output or generated files",
+                "a concise summary of what was processed",
+            ],
+            "escalate_if": [
+                "the script needs significant algorithm or architecture design",
+                "requirements are ambiguous before coding can start",
+            ],
+            "request_preview": preview,
+        }
+
+    if route == "high_compute_model":
+        return {
+            "execution_type": "run_high_compute_model",
+            "goal": "Handle the request with deeper reasoning before taking action.",
+            "immediate_action": "Use a stronger model to analyze the task, resolve ambiguity, and produce the answer or plan.",
+            "codex_instruction": "Give the task to a stronger model path first. If execution is later needed, convert the resulting plan into code or commands.",
+            "artifacts_to_produce": [
+                "a detailed answer, design, or plan",
+                "explicit tradeoffs, assumptions, or decision criteria",
+            ],
+            "escalate_if": [
+                "the task becomes procedural after planning",
+                "the answer requires file processing or repeatable transformations",
+            ],
+            "request_preview": preview,
+        }
+
+    return {
+        "execution_type": "run_low_compute_model",
+        "goal": "Handle the request with the cheapest viable language-model pass.",
+        "immediate_action": "Use a lightweight model path for a fast first answer.",
+        "codex_instruction": "Start with a cheaper/faster model. Escalate only if the output is weak, incomplete, or the task expands.",
+        "artifacts_to_produce": [
+            "a short answer or rewrite",
+            "minimal reasoning with quick turnaround",
+        ],
+        "escalate_if": [
+            "the request turns out to be ambiguous",
+            "the first pass fails quality checks",
+            "multiple retries would cost more than escalating once",
+        ],
+        "request_preview": preview,
+    }
+
+
+def normalize(text: str) -> str:
+    text = text.strip().lower()
+    text = re.sub(r"\s+", " ", text)
+    return text
+
+
+def keyword_hits(text: str, keywords: List[str]) -> List[str]:
+    hits = []
+    for keyword in keywords:
+        if keyword in text:
+            hits.append(keyword)
+    return hits
+
+
+def score_python_route(text: str) -> RouteScore:
+    reasons: List[str] = []
+    score = 0
+
+    deterministic_hits = keyword_hits(
+        text,
+        [
+            "python",
+            "script",
+            "csv",
+            "json",
+            "yaml",
+            "xml",
+            "excel",
+            "spreadsheet",
+            "parse",
+            "extract",
+            "transform",
+            "convert",
+            "rename",
+            "batch",
+            "directory",
+            "folder",
+            "file",
+            "files",
+            "dataset",
+            "log",
+            "logs",
+            "calculate",
+            "count",
+            "sort",
+            "filter",
+            "regex",
+            "scrape",
+        ],
+    )
+    if deterministic_hits:
+        score += 4 + min(len(deterministic_hits), 6)
+        reasons.append(
+            "deterministic data/file-processing signals: "
+            + ", ".join(deterministic_hits[:6])
+        )
+
+    if any(token in text for token in ["automate", "repeatedly", "pipeline", "generate report"]):
+        score += 3
+        reasons.append("request looks repetitive or automation-friendly")
+
+    if any(token in text for token in ["exact", "precise", "reproducible", "structured output"]):
+        score += 2
+        reasons.append("request favors reproducible execution over free-form reasoning")
+
+    return RouteScore("python_script", score, reasons)
+
+
+def score_high_route(text: str) -> RouteScore:
+    reasons: List[str] = []
+    score = 0
+
+    reasoning_hits = keyword_hits(
+        text,
+        [
+            "analyze",
+            "analysis",
+            "design",
+            "architect",
+            "strategy",
+            "compare",
+            "tradeoff",
+            "debug",
+            "root cause",
+            "plan",
+            "complex",
+            "hard",
+            "unclear",
+            "ambiguous",
+            "research",
+            "brainstorm",
+            "proposal",
+            "spec",
+        ],
+    )
+    if reasoning_hits:
+        score += 4 + min(len(reasoning_hits), 6)
+        reasons.append(
+            "open-ended reasoning signals: " + ", ".join(reasoning_hits[:6])
+        )
+
+    if any(
+        token in text
+        for token in ["step by step", "carefully", "deeply", "thoroughly", "rigorous"]
+    ):
+        score += 3
+        reasons.append("user explicitly asks for deeper or more careful reasoning")
+
+    if len(text.split()) > 80:
+        score += 2
+        reasons.append("request is long enough to suggest higher-context reasoning")
+
+    return RouteScore("high_compute_model", score, reasons)
+
+
+def score_low_route(text: str) -> RouteScore:
+    reasons: List[str] = []
+    score = 0
+
+    lightweight_hits = keyword_hits(
+        text,
+        [
+            "rewrite",
+            "rephrase",
+            "translate",
+            "summarize",
+            "summary",
+            "classify",
+            "tag",
+            "format",
+            "clean up",
+            "fix grammar",
+            "short answer",
+            "quick",
+            "simple",
+        ],
+    )
+    if lightweight_hits:
+        score += 4 + min(len(lightweight_hits), 5)
+        reasons.append(
+            "lightweight language-task signals: " + ", ".join(lightweight_hits[:6])
+        )
+
+    if len(text.split()) <= 25:
+        score += 2
+        reasons.append("request is short and likely cheap to answer")
+
+    if any(token in text for token in ["cheap", "fast", "brief"]):
+        score += 2
+        reasons.append("user is optimizing for speed or lower cost")
+
+    return RouteScore("low_compute_model", score, reasons)
+
+
+def choose_route(text: str) -> Dict[str, object]:
+    normalized = normalize(text)
+    if not normalized:
+        execution_plan = build_execution_plan("low_compute_model", text, 0.25)
+        return {
+            "route": "low_compute_model",
+            "confidence": 0.25,
+            "reasons": ["empty request defaults to the lowest-cost model"],
+            "scores": {route: 0 for route in ROUTES},
+            "execution_plan": execution_plan,
+        }
+
+    scored_routes = [
+        score_python_route(normalized),
+        score_high_route(normalized),
+        score_low_route(normalized),
+    ]
+    scored_routes.sort(key=lambda item: item.score, reverse=True)
+
+    winner = scored_routes[0]
+    runner_up = scored_routes[1]
+
+    if winner.score == 0:
+        winner = RouteScore(
+            "high_compute_model",
+            1,
+            ["fallback to the stronger model because the task is not obviously deterministic or trivial"],
+        )
+        runner_up = RouteScore("low_compute_model", 0, [])
+
+    margin = max(winner.score - runner_up.score, 0)
+    confidence = min(0.55 + 0.1 * margin, 0.95)
+
+    recommended_next_action = {
+        "python_script": "Prefer executing or writing a Python script first, then use a model only for glue logic or explanation.",
+        "high_compute_model": "Prefer a stronger model for planning, ambiguity resolution, or multi-step reasoning.",
+        "low_compute_model": "Prefer a cheaper/faster model for the first pass and escalate only if it struggles.",
+    }[winner.name]
+    confidence = round(confidence, 2)
+    execution_plan = build_execution_plan(winner.name, text, confidence)
+
+    return {
+        "route": winner.name,
+        "confidence": confidence,
+        "reasons": winner.reasons,
+        "scores": {item.name: item.score for item in scored_routes},
+        "recommended_next_action": recommended_next_action,
+        "execution_plan": execution_plan,
+    }
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description="Route a request to python_script, high_compute_model, or low_compute_model."
+    )
+    parser.add_argument("--text", help="Request text to classify. If omitted, read from stdin.")
+    parser.add_argument(
+        "--pretty",
+        action="store_true",
+        help="Pretty-print JSON output.",
+    )
+    parser.add_argument(
+        "--summary",
+        action="store_true",
+        help="Print a compact human-readable routing summary instead of JSON.",
+    )
+    args = parser.parse_args()
+
+    text = args.text if args.text is not None else sys.stdin.read()
+    result = choose_route(text)
+    if args.summary:
+        print(f"Route: {result['route']}")
+        print("Why: " + "; ".join(result["reasons"][:2]))
+        print("Next step: " + result["execution_plan"]["immediate_action"])
+    elif args.pretty:
+        print(json.dumps(result, indent=2, ensure_ascii=True))
+    else:
+        print(json.dumps(result, ensure_ascii=True))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/skills/task-router/test_config.json b/skills/task-router/test_config.json
new file mode 100644
index 0000000..44c70d8
--- /dev/null
+++ b/skills/task-router/test_config.json
@@ -0,0 +1,15 @@
+{
+  "component_type": "dialogue_reading",
+  "format": {
+    "text": "string",
+    "audio": "optional_file",
+    "language": "string"
+  },
+  "config": {
+    "duration": 30,
+    "scoring_threshold": 80,
+    "text": "Hello, how are you?",
+    "language": "en"
+  },
+  "status": "generated"
+}
\ No newline at end of file