菜单
Model string conventions for set_agent_model — how to format names for Gemini, Claude, OpenRouter, and other providers. Load this any time the user asks to change, swap, set, switch, route, or test a model — to avoid building the wrong model string.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/misunders2d/ori --skill model-swap-skill命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | model-swap-skill |
| description | Model string conventions for set_agent_model — how to format names for Gemini, Claude, OpenRouter, and other providers. Load this any time the user asks to change, swap, set, switch, route, or test a model — to avoid building the wrong model string. |
set_agent_model and verify_model_reachable take a model string in <provider-prefix>/<rest> form. Get the prefix right or the probe rejects with UNKNOWN_PROVIDER.
| Prefix | When to use | Example |
|---|---|---|
litellm/ | LiteLlm-routed (default for hot-swap). Any model litellm supports — Gemini, Anthropic, OpenAI, Mistral, Cohere, Groq, etc. | litellm/gemini/gemini-3.1-flash-lite-preview |
gemini/ | Native ADK Gemini class. Required for pinned components (google_search, embedding). Same-provider swaps only. | gemini/gemini-3.1-flash-lite-preview |
anthropic/ | Native ADK Claude class. | anthropic/claude-sonnet-4-6 |
openrouter/ | OpenRouter (multi-provider routing through one API). Goes through LiteLlm under the hood. | openrouter/deepseek/deepseek-v4-flash |
If you see UNKNOWN_PROVIDER in the response, you used a vendor name as the prefix instead of one of the four above.
OpenRouter model IDs on https://openrouter.ai/models look like vendor/model-name (e.g. deepseek/deepseek-v4-flash, anthropic/claude-3.5-sonnet, meta-llama/llama-3.1-405b-instruct). To use them in Ori:
openrouter/<vendor>/<model>
The openrouter/ prefix is REQUIRED. It tells Ori's PROVIDER_REGISTRY to route through OpenRouter; without it, the probe rejects the model.
Examples:
| User says | Correct model string |
|---|---|
| "deepseek-v4-flash via openrouter" | openrouter/deepseek/deepseek-v4-flash |
| "claude sonnet via openrouter" | openrouter/anthropic/claude-3.5-sonnet |
| "gpt-4o via openrouter" | openrouter/openai/gpt-4o |
| "llama 405b via openrouter" | openrouter/meta-llama/llama-3.1-405b-instruct |
| "grok via openrouter" | openrouter/x-ai/grok-2 |
WRONG (probe will reject): deepseek/deepseek-v4-flash, openai/gpt-4o, x-ai/grok-2 — these have no registered provider prefix.
If the user just says "use Claude Sonnet 4.6" (no OpenRouter mentioned and they have an ANTHROPIC_API_KEY configured), prefer native or LiteLlm-routed:
litellm/anthropic/claude-sonnet-4-6 — hot-swappable, recommendedanthropic/claude-sonnet-4-6 — native Claude class, locked at construction (can't cross-provider swap later)Default to litellm/... unless the user specifically asks for the native class.
google_search and embedding are PINNED to native classes (search grounding + embedding endpoint require it). Attempts to swap them return COMPONENT_PINNED. Don't try.
list_available_models to see them.openrouter/.verify_model_reachable("<model>") first to probe without persisting.set_agent_model("<Component>", "<model>") — runs the probe, persists on success.| User intent | Build |
|---|---|
| "switch coordinator to deepseek-v4-flash via openrouter" | set_agent_model("CoordinatorAgent", "openrouter/deepseek/deepseek-v4-flash") |
| "use claude for the developer" | set_agent_model("DeveloperAgent", "litellm/anthropic/claude-sonnet-4-6") |
| "test if openrouter/x-ai/grok-2 works" | verify_model_reachable("openrouter/x-ai/grok-2") |
| "reset everyone to defaults" | reset_model_override("") (empty = all) |
| "what model is the coordinator on?" | list_available_models() and read the row |