# Model Switch

> Switch the Manager Agent's own LLM model. Use when the human admin requests changing the Manager's model.

- Skill: `agentscope-ai/model-switch-2` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx get-skill add agentscope-ai/model-switch-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/agentscope-ai/model-switch-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: agentscope-ai (https://skillmd.com/u/agentscope-ai)
- Updated: 2026-11-15
- Page: https://skillmd.com/skills/agentscope-ai/model-switch-2

---


# Model Switch

Switch the Manager's own LLM model. The script tests connectivity first, then patches `openclaw.json`.

## Usage

```bash
bash /opt/agentteams/agent/skills/model-switch/scripts/update-manager-model.sh <MODEL_ID> [--context-window <SIZE>] [--no-reasoning]
```

Examples:
```bash
bash /opt/agentteams/agent/skills/model-switch/scripts/update-manager-model.sh claude-sonnet-4-6
bash /opt/agentteams/agent/skills/model-switch/scripts/update-manager-model.sh my-custom-model --context-window 300000
bash /opt/agentteams/agent/skills/model-switch/scripts/update-manager-model.sh deepseek-chat --no-reasoning
```

## What the script does

1. Strips any `agentteams-gateway/` prefix from the model name
2. Tests the model via `POST /v1/chat/completions` on the AI Gateway — exits with error if unreachable
3. If the model is already in the `models` array: refreshes its `reasoning` / `contextWindow` / `maxTokens` / `input` fields and switches `agents.defaults.model.primary`
4. If the model is new: adds it to the `models` array and switches primary
5. Outputs the apply/verification instructions for your Manager runtime

Because step 3 rewrites the model's fields, re-running the script is the way to
correct an entry that was first written with wrong values — for example a model
that was added with the 150,000 default because nobody knew its real window:
run the script again with `--context-window <SIZE>` and the new value lands in
`openclaw.json`.

## After running the script

Follow the instructions for your `AGENTTEAMS_MANAGER_RUNTIME`:

- **OpenClaw**: run `openclaw gateway restart`, then verify the active model.
- **QwenPaw**: the script updates the standard `openclaw.json` consumed by your existing background bridge. Wait for its next watcher cycle (normally within 60 seconds). Read back `active_llm.provider_id` and `active_llm.model` from your native secret `providers.json` (honor `QWENPAW_SECRET_DIR`), and verify your next request uses the requested model before reporting completion. A `CONFIG_UPDATED` message alone confirms the standard file was written. If the native state stays stale, inspect your bridge logs and report the pending apply step.
- **Legacy CoPaw**: follow the `RESTART_REQUIRED` instruction to restart your CoPaw service.

For QwenPaw, verify context and reasoning settings in your native runtime separately; writing these fields to the standard config does not prove every option is supported or applied.

## Reasoning control

By default, reasoning (extended thinking) is enabled. To disable it, pass `--no-reasoning`.

## On failure

If the gateway test fails (non-200), the script outputs `ERROR: MODEL_NOT_REACHABLE` and exits. No changes are made to `openclaw.json`.

When you see this error, tell the human admin clearly:

1. The model is not reachable because the current default AI Provider likely does not support it.
2. They need to open the Higress Console and:
   - Create a **new AI Provider** for the model's vendor (e.g. `kimi`, `deepseek`, `minimax`).
   - Create a **new AI Route** with a model name prefix predicate (e.g. provider `kimi` → match `kimi-*`), so that requests for models with that prefix are routed to the new provider, while unmatched models continue to go through the default route.
3. **Do NOT modify the default AI Provider** — it is managed by the initialization config and will be overwritten on restart.

After the admin confirms the provider and route are configured, you can retry the model-switch script.

## Important

You switch your **primary model** persistently through the standard config. Follow your runtime-specific apply and verification steps above. With OpenClaw, the human admin can also use `/model` to switch the current session instantly; that does not persist the primary model and only supports pre-configured models.

## Switching to an unknown model

When the human admin requests switching to a model you don't recognize, you MUST:

1. **Ask the admin two questions** before running the script:
   - "This model is not in the known list. What is its context window size (in tokens)?"
   - "Does this model support reasoning (extended thinking)?"
2. Run the script with the appropriate flags:
   ```bash
   bash /opt/agentteams/agent/skills/model-switch/scripts/update-manager-model.sh <MODEL_ID> --context-window <SIZE> [--no-reasoning]
   ```
3. If the admin does not know the context window, use the default (150,000) by omitting `--context-window`.
4. If the model does not support reasoning, add `--no-reasoning`.

## Pre-configured models (for reference)

| Model | contextWindow | maxTokens |
|-------|--------------|-----------|
| gpt-5.4 | 1,050,000 | 128,000 |
| gpt-5.3-codex / gpt-5-mini / gpt-5-nano | 400,000 | 128,000 |
| claude-opus-4-6 | 1,000,000 | 128,000 |
| claude-sonnet-4-6 | 1,000,000 | 64,000 |
| claude-haiku-4-5 | 200,000 | 64,000 |
| qwen3.5-plus | 200,000 | 64,000 |
| deepseek-chat / deepseek-reasoner / kimi-k2.5 | 256,000 | 128,000 |
| glm-5 / MiniMax-M2.7 / MiniMax-M2.7-highspeed / MiniMax-M2.5 | 200,000 | 128,000 |

