メニュー
Dynamic LLM model management. Switch between model tiers (flagship, worker, efficiency, reasoning, coding) at runtime to optimize for intelligence, speed, or cost.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Record audio from the Pi's USB microphone or play WAV files through the Pi's speaker. The audio subsystem runs on the host (audio bridge on port 18796); this skill triggers record/play from inside the container. Use for "record the mic", "play a sound", or when the user wants to leave a voice note (e.g. after an RFID scan with voice announce).
Take a still photo from the USB camera connected to the Pi. Use when the user asks to take a photo, snap a picture, capture from the camera, or get an image from the webcam. The camera runs on the Pi host; this skill triggers the host capture service and then sends the image to Telegram.
Fashion trend intelligence engine. Scans editorial sites and social media for current trends, maintains persistent knowledge, and delivers curated trend reports personalized to the customer's style profile.
Replay learned IR button presses (e.g. fan power, fan speed) by remote name and button name. Use when the user wants to control an IR device (fan, TV, etc.) and buttons were previously learned with scripts/ir-learn.py on the Pi. Only applies when the IR dongle is available.
Read PDFs in workspace: extract text for normal PDFs, or convert pages to images and use vision for scanned/image-heavy PDFs. Use when the user asks to read a PDF, summarize a document, convert a PDF to images, or answer questions about a PDF.
High-Density Physical Admin Dashboard (16x2 LCD). Use this skill to update the physical display on the Pi.
SOC 職業分類に基づく
| name | llm-manager |
| description | Dynamic LLM model management. Switch between model tiers (flagship, worker, efficiency, reasoning, coding) at runtime to optimize for intelligence, speed, or cost. |
| metadata | {"openclaw":{"emoji":"⚙️","requires":{"bins":["bash","python3"]}}} |
You are an expert model orchestrator. You can swap your own underlying LLM provider and tier at runtime to adapt to the complexity of your current task.
gpt-5.1 / gpt-5): Use for high-level reasoning, complex planning, and critical decision-making. High cost, high intelligence.gpt-5-mini): The Sweet Spot. Excellent for most general tasks, web browsing, and tool execution. Balanced cost and intelligence.gpt-5-nano): Ultra-low cost. Use for reports, logging, simple summarization, or low-importance heartbeats.o3 / o4-mini): Use for deep logic, complex math, or resolving ambiguity when other models struggle.gpt-5-codex): Optimized for software engineering and heavy script generation.Update your primary model or a specific task model (heartbeat/report) to a new tier.
Command:
exec: bash /home/openclaw/.openclaw/workspace/skills/llm-manager/scripts/switch-tier.sh <tier> [task_type] "[reason]"
flagship, worker, efficiency, reasoning, codingprimary (default), heartbeat, reportExample:
exec: bash /home/openclaw/.openclaw/workspace/skills/llm-manager/scripts/switch-tier.sh flagship primary "Upgrading for complex architectural analysis"
Example (downgrading heartbeats to save cost):
exec: bash /home/openclaw/.openclaw/workspace/skills/llm-manager/scripts/switch-tier.sh efficiency heartbeat "Saving cost on idle checks"
All model switches are logged to /tmp/openclaw/model_switches.log and will appear in the IronClaw Daily Report.
Note: Model switches take effect immediately via config hot-reload. They are session-scoped — switches reset when the agent is restarted via compose-up.sh. If a switch proves valuable long-term, notify the owner so they can make it permanent in config/openclaw.json.