task-based-model-router（基于任务的模型路由器）

Pre-flight Check: PaleBlueDot AI TokenRouter Provider Configuration

This skill requires a PaleBlueDot AI TokenRouter provider to be configured. Before using ANY feature, the agent MUST perform the following check-and-setup flow.

Phase 1: Check Existing Provider

1. Read the configuration file at ~/.openclaw/openclaw.json.
2. Look under models.providers for any provider whose baseUrl contains https://open.palebluedot.ai (partial match is sufficient).
3. If found → Skip to Phase 3 (the provider is already configured).
4. If NOT found → Proceed to Phase 2.

Phase 2: Guide User & Write Configuration

When no TokenRouter provider is detected, the agent should:

1. Inform the user that TokenRouter is not configured and prompt them to register:

> TokenRouter provider not configured. To use this skill, you need a PaleBlueDot AI TokenRouter account. > Please visit https://www.palebluedot.ai to register and log in, then go to the TokenRouter section to get your configuration (base URL and API key). > Once you have the information, please provide it to me and I will complete the configuration for you.

1. Wait for the user to provide the following information:

- Base URL (e.g., https://open.palebluedot.ai/v1) - API Key (the user's real API key from PaleBlueDot AI) - *(Optional)* Provider name (defaults to tokenrouter if not specified)

1. Agent writes the configuration by running the setup command:

   setup --name <provider_name> --base-url <baseUrl> --api-key <apiKey>

This writes the provider into ~/.openclaw/openclaw.json under models.providers with a backup of the original config.

1. Do NOT use placeholder API keys. The agent must only write the actual credentials provided by the user.

Phase 3: Re-verify & Auto-Sync Models

After the provider is confirmed (either pre-existing or just written in Phase 2), the agent MUST:

1. Re-verify the provider configuration by running:

   check

This confirms the provider with baseUrl containing https://open.palebluedot.ai is properly saved in the config.

1. If verification fails, inform the user and go back to Phase 2.

1. If verification succeeds, automatically sync models by running:

   sync

This will: - Fetch all available models from the TokenRouter API. - Add all model names to the TokenRouter provider's models array. - Add all models to the models.allowed list. - All models will be routed through the configured TokenRouter provider. - Display the synced model list to the user.

1. Confirm completion to the user:

> TokenRouter configuration complete. {N} models have been synced and added to your allow list. All models are routed through the TokenRouter provider. You can now use list to view pricing or start planning tasks.

Automatic Pre-Planning Sync

Before executing any plan command, the system will automatically:

1. Verify TokenRouter Provider: Checks if a provider whose baseUrl contains https://open.palebluedot.ai exists in models.providers. If not found, the agent enters the Phase 2 setup flow described above.
2. Auto-Sync Models: Fetches the latest model list and updates the provider's models array and the models.allowed list.
3. Set Default Model: If no default model is set, picks the first available model from the synced list.

This ensures that the plan command always has access to the most up-to-date model information, and all models are routed through the user's TokenRouter provider.

Dynamic Model Tier System

Model tiers are not hardcoded. On every plan invocation, the system:

1. Fetches the live price list from TokenRouter API.
2. Filters to well-known models (GPT, Claude, Gemini, DeepSeek, Llama, Qwen, Grok).
3. Computes each model's output price.
4. Sorts by price descending and splits into 3 equal buckets (high / mid / low).
5. Picks the median-priced model from each bucket to avoid outliers.

If the API is unreachable or fewer than 3 known models are available, the system falls back to hardcoded defaults (claude-opus-4.6 / gpt-4o-mini / deepseek-v3.2).

6-Category Task Classification Engine

The plan command uses an enhanced classifier that supports both Chinese and English keywords. Tasks are scored against 6 categories; the highest-scoring category wins.

Categories & Routing Pipelines

1. Coding

Keywords: code, program, script, debug, function, api, database, app, test, bug, deploy, refactor, 代码, 编程, 脚本, 程序, 调试, 测试, 开发, 接口, 部署, 重构, 修复, 函数, 算法, 数据库, 前端, 后端, 全栈, 爬虫, 框架, 模块 ...

2. Analysis

Keywords: analyze, compare, evaluate, research, report, data, statistics, 分析, 对比, 评估, 研究, 调研, 报告, 数据, 统计, 洞察, 指标, 复盘 ...

3. Writing

Keywords: write, article, blog, content, story, email, essay, documentation, 写作, 文章, 博客, 内容, 故事, 邮件, 文案, 文档, 稿件, 撰写, 起草 ...

4. Creative

Keywords: creative, brainstorm, idea, design, logo, prototype, 创意, 头脑风暴, 点子, 设计, 艺术, 原型, 线框图, 灵感, 构思 ...

5. Translation

Keywords: translate, localize, interpretation, 翻译, 本地化, 国际化, 多语言, 中译英, 英译中 ...

6. Simple

Keywords: simple, quick, summarize, list, count, define, explain, 简单, 快速, 总结, 列出, 计数, 查询, 定义, 解释, 概括 ...

Cost Savings by Category

Multi-Agent Routing Guidance

When the user's task is complex (e.g., building an application, designing a system, multi-step workflows), proactively suggest multi-agent routing:

1. Identify complexity — If the task involves multiple phases, recommend running plan.
2. Show the pricing — Run list to display the full price list.
3. Run the planner — Run plan "<task>" to show the recommended routing with projected savings.
4. Offer to enable models — Ask the user if they want to enable the recommended models as fallbacks.

Example prompt to user:

Your task involves multiple phases. Let me run the planner to find the optimal routing: `` build a REST API with authentication / 帮我构建一个REST API的鉴权功能 `` The planner will automatically pick the best models for each phase based on current pricing and show projected savings.

Quick Start

# Step 1: Check if TokenRouter is configured
check

# Step 2: If not configured, set it up (agent does this with user-provided credentials)
setup --name tokenrouter --base-url https://open.palebluedot.ai/v1 --api-key sk-xxx...

# Step 3: Verify and sync all models
check
sync

# List all models with real-time pricing
list / 列出TokenRouter的模型价格列表

# Get routing recommendations for a task (Chinese or English)
# NOTE: This will automatically sync models before planning
write a Python script
帮我开发一个用户管理后端接口
analyze and compare the data reports of three competing products
把这段中译英

# Enable a model by index or name
enable 1
enable openai/gpt-4o-mini

# Generate execution plan for host agent to dispatch sub-agents
build a todo app

Core Functions

1. check - Verify Provider Configuration

check / 检查TokenRouter配置

Verifies that a TokenRouter provider (baseUrl containing https://open.palebluedot.ai) exists in the config. Displays provider details if found, or guides the user to set up if not.

2. setup - Write Provider Configuration

setup --name tokenrouter --base-url https://open.palebluedot.ai/v1 --api-key <key>

Writes the TokenRouter provider into ~/.openclaw/openclaw.json (with automatic backup). The agent uses this command after the user provides their credentials from PaleBlueDot AI. The agent must never use placeholder API keys — only real credentials provided by the user.

3. sync - Fetch & Sync All Models

sync / 同步TokenRouter模型

Fetches all available models from the TokenRouter API, adds them to the provider's models array and the models.allowed list. All models are routed through the configured TokenRouter provider. This is automatically called before plan.

4. list - Real-Time Model Pricing

list / 列出模型的价格列表

Fetches current TokenRouter pricing and displays all available models with input/output/cache prices.

5. plan - Smart Task Routing

plan "<task description>" / 计划 "<任务描述>"
plan "<task description>" --execute / 计划 "<任务描述>" 并执行

Classifies the task, builds a multi-phase pipeline, assigns dynamic model tiers, and shows projected savings. Automatically syncs models before planning. Add --execute (or -x) to generate a structured JSON execution plan (swarm_plan.json) that the host agent uses to dispatch sub-agents via the internal sessions_spawn API.

How --execute works:

The --execute flag does NOT call sub-agents directly via CLI. Instead, it outputs a JSON plan to ~/.openclaw/workspace/swarm_plan.json and to stdout. The host agent (OpenClaw) reads this plan and dispatches sub-agents internally using sessions_spawn. Each step in the plan contains:

• model: which model to use for this phase
• system_prompt: the role prompt for the sub-agent
• task_prompt: the task instruction
• expected_artifact: the file the sub-agent should produce
• max_retries and timeout_seconds: retry and timeout policy

Steps must be executed sequentially — each step's artifact is context for the next.

6. enable - Auto-Configuration

enable 1                    # Enable model by index
enable openai/gpt-4o-mini   # Enable by name
启用 1
启用 openai/gpt-4o-mini

Writes the model directly into ~/.openclaw/openclaw.json (with automatic backup).

Adaptive Stability Fallback

The system tracks historical model performance via swarm_memory.json and swarm_insights.json:

• Tier selection: If a tier2/tier3 model's historical success rate drops below 50%, it is automatically replaced by the tier1 model for that phase during plan generation.
• Retry policy: The generated execution plan specifies max_retries: 2 per step. The host agent should inject the error context into the retry prompt when re-dispatching a failed step.
• Logging: The host agent should append execution results to swarm_memory.json. Run consolidate_memory.py to generate performance insights for future adaptation.

Advanced Usage

Custom Routing Rules

You can override the dynamic tier selection for specific categories by creating ~/.openclaw/model-routing.json. Each category maps to an ordered list of model IDs, one per phase (matching the pipeline order). An optional "fallback" model is used for any phase without an explicit override.

// ~/.openclaw/model-routing.json
{
  "coding": [
    "anthropic/claude-opus-4.6",
    "openai/gpt-4o-mini",
    "deepseek/deepseek-v3.2"
  ],
  "analysis": [
    "google/gemini-3-pro-preview",
    "anthropic/claude-sonnet-4.6",
    "deepseek/deepseek-v3.2"
  ],
  "writing": [
    "anthropic/claude-opus-4.6",
    "openai/gpt-4o",
    "openai/gpt-4o-mini"
  ],
  "creative": [
    "anthropic/claude-opus-4.6",
    "openai/gpt-4o"
  ],
  "translation": [
    "openai/gpt-4o",
    "deepseek/deepseek-v3.2"
  ],
  "simple": [
    "deepseek/deepseek-v3.2"
  ],
  "fallback": "openai/gpt-4o-mini"
}

How it works:

• The list index corresponds to the phase order in each category's pipeline (e.g., for coding: index 0 = Design, 1 = Code, 2 = Review).
• If a category is not listed, the default dynamic tier selection applies.
• If the list is shorter than the number of phases, remaining phases use "fallback" (if set) or the default tier.
• If the file does not exist, the system uses fully dynamic tier selection.