Friday Router

Austin's intelligent model routing skill with fixed bugs from the original intelligent-router and customized for his preferred models.

Default model: OpenRouter Claude Sonnet 4 — capable of basic coding and complex tasks. The router down-routes to cheaper/faster models (e.g. Gemini 2.5 Flash) only when the task is clearly simple (FAST tier). So you start capable by default; simple tasks save cost.

Orchestrator flow (mandatory for task requests)

When you are the main agent (Sonnet) and the user gives you a task (something to do: code, research, check, build, fix, write, create, design, poem, story, etc.), you must delegate it. You do not perform the task yourself; a sub-agent does.

Steps (both required):

1. Run the Friday Router to get spawn params:

   python3 workspace/skills/friday-router/scripts/router.py spawn --json "<user message>"

   The --json output is a single JSON object with model, task, sessionTarget, and optionally label.

2. Call sessions_spawn with the task and model (and any other params the tool accepts) from the router output. Use exactly the model value from the router—do not substitute a different model. The sub-agent will run with that model (e.g. Kimi k2.5 for CREATIVE) and do the work.

3. Forward or summarize the sub-agent's result to the user. You may briefly say you're delegating to the recommended model, then present the sub-agent's reply.

Anti-pattern (forbidden): Running the router, seeing e.g. "Kimi k2.5 recommended", and then doing the creative task yourself in the same turn. That bypasses the sub-agent: Kimi is never used and logs will show only the main model (e.g. Claude Sonnet 4). Always call sessions_spawn after the router so the recommended model actually runs the task.

Exception: Meta-questions (e.g. "what model are you?", "how does routing work?") you answer yourself. Only delegate when the user is asking for work to be done.

What Changed from Original

Bug	Fix
Simple indicators inverted (high match = complex)	Now correctly: high simple keyword match = FAST tier
Agentic tasks not bumping tier	Multi-step tasks now properly bump to CODE tier
Vision tasks misclassified	Vision keywords now take priority over other classifications
Code keywords not detected	Added React, JWT, API, and other common code terms
Confidence always low	Now varies appropriately based on keyword match strength

Model Selection (Austin's Prefs)

Use Case	Primary	Fallback
Default (session)	OpenRouter Claude Sonnet 4	—
Fast/F cheap	Gemini 2.5 Flash (OpenRouter)	Gemini 1.5 Flash, Haiku
Reasoning	GLM-5	Minimax 2.5
Creative/Frontend	Kimi k2.5	—
Research	Grok Fast	—
Code/Engineering	DeepSeek-Coder-V2	Qwen2.5-Coder
Quality/Complex	OpenRouter Claude Sonnet 4	Claude 3.5 Sonnet, GPT-4o
Vision/Images	GPT-4o	—

When the skill is installed, OpenClaw's default model is set to OpenRouter Claude Sonnet 4 so every new session is capable by default. The router then recommends Gemini 2.5 Flash only for clearly simple tasks (check, status, list, etc.).

Usage

CLI

# Show session default model (capable by default)
python scripts/router.py default

# Classify a task
python scripts/router.py classify "fix lint errors in utils.js"

# Show detailed scoring
python scripts/router.py score "build a React auth system"

# Estimate cost
python scripts/router.py cost "design a landing page"

# Get OpenClaw spawn params (human-readable)
python scripts/router.py spawn "research the best LLMs"

# Get spawn params as JSON (for sessions_spawn: parse and pass model, task, etc.)
python scripts/router.py spawn --json "research the best LLMs"

# List all models
python scripts/router.py models

In Code

from scripts.router import FridayRouter

router = FridayRouter()

# Classify task
tier = router.classify_task("check server status")
# → "FAST"

# Get model recommendation
result = router.recommend_model("build authentication system")
# → {tier: "CODE", model: {...}, fallback: {...}}

# Spawn with correct model
spawn_params = router.spawn_agent("fix this bug", label="bugfix")
# → {params: {model: "deepseek-ai/deepseek-coder-v2", ...}}

Tier Detection

• FAST: check, get, list, show, status, monitor, fetch, simple
• REASONING: prove, logic, analyze, derive, math, step by step
• CREATIVE: creative, write, story, design, UI, UX, frontend, website
• RESEARCH: research, find, search, lookup, web, information
• CODE: code, function, debug, fix, implement, refactor, test, React, JWT
• QUALITY: complex, architecture, design, system, comprehensive
• VISION: image, picture, photo, screenshot, visual

Examples

# Simple monitoring → FAST
python scripts/router.py classify "check server status"
# → FAST (Gemini 2.5 Flash)

# Code task → CODE
python scripts/router.py classify "fix lint errors in utils.js"
# → CODE (DeepSeek Coder)

# Multi-step build → CODE (bumped from default)
python scripts/router.py classify "build React auth with JWT, tests, CI"
# → CODE (DeepSeek Coder) - agentic task detected

# Mathematical proof → REASONING
python scripts/router.py classify "prove sqrt(2) is irrational"
# → REASONING (GLM-5)

# Design task → CREATIVE
python scripts/router.py classify "design a beautiful landing page"
# → CREATIVE (Kimi k2.5)

# Research → RESEARCH
python scripts/router.py classify "find information about GPT-5"
# → RESEARCH (Grok Fast)

# Vision → VISION (priority over other keywords)
python scripts/router.py classify "analyze this screenshot"
# → VISION (GPT-4o)