model-swap-skill

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.

Model Swap — String Format

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.

Registered providers

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 format — the most common trap

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.

Same-provider via LiteLlm vs native

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, recommended
• anthropic/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.

Pinned components

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.

Procedure

1. Parse the user's intent: which component (CoordinatorAgent / DeveloperAgent / KnowledgeAgent / summarizer) and which provider/model.
2. If unsure, ASK — don't guess. The list of components: call list_available_models to see them.
3. Construct the model string per the prefix table above. If the user said "via openrouter", the string MUST start with openrouter/.
4. (Optional) Call verify_model_reachable("<model>") first to probe without persisting.
5. Call set_agent_model("<Component>", "<model>") — runs the probe, persists on success.
6. Confirm to the user with the result. If the probe failed, relay the error verbatim.

What the user said vs what to build

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