hudl-model-switch（hudl 模型切换）

hudl-model-switch

Switch the active LLM model for this OpenClaw agent via the Huddle01 GRU gateway.

Non-negotiable rules

• This skill only applies to the hudl provider backed by gru.huddle01.io.
• The target model value in config must always be hudl/<model-id>.
• Never edit JSON manually for model switching. Always use scripts/switch-model.sh.
• Never modify provider wiring during a switch (models.providers.hudl.baseUrl, apiKey, and other unrelated keys stay unchanged).
• Treat any post-switch mismatch, missing model field, or non-hudl/ result as a hard failure. Do not restart in that state.
• For status/reporting, the active agent model is the source of truth. agents.defaults.model.primary is only a default and may be stale.

Prerequisites

This skill only works when the OpenClaw config has a hudl provider pointing at gru.huddle01.io. If the provider is missing or the baseUrl is different, stop and tell the user this skill requires the Huddle01 GRU gateway.

Step-by-step

On any switch/change request

Step 1: Validate provider + config path (required)

Run:

bash <skill_dir>/scripts/validate.sh

• Exit 0: continue.
• Exit 1: stop. Show the exact validation error and tell the user no config changes were made.
• Validation resolves config in this order: OPENCLAW_CONFIG, ~/.openclaw/config.json, ~/.openclaw/openclaw.json.
• The provider check is strict: the hudl provider must exist, apiKey must be present, and baseUrl must point at host gru.huddle01.io.

Step 2: Resolve the exact target model ID (required)

Use <skill_dir>/references/models.md to map user intent to a canonical model.

Normalization rules:

• If the resolved model does not start with hudl/, prepend hudl/.
• If the user already supplied hudl/..., keep it unchanged.
• Never write a non-hudl/ model into config.
• Reject empty model IDs, hudl/ by itself, or values with spaces.
• For unknown models, still normalize to hudl/<user-value> and warn clearly that the model must exist on GRU or requests will fail.

Ambiguity rules:

• If request is ambiguous (for example "claude" could be Opus/Sonnet/Haiku), choose the catalog default alias mapping (currently hudl/claude-opus-4.6) and state what was chosen.
• If the user asks for a specific variant, obey that exact variant.

Step 3: Capture current state before mutation (required)

Read and note current values so the response can show before/after:

• active agent model (agents.list[*].model.primary, preferring agent id: "main")
• default model (agents.defaults.model.primary)
• If .agents.list has no entries, stop. This skill is for switching an active agent model, so a config with no agent entries is not a valid target.
• When active and default differ, describe that as a mismatch and treat the active agent model as the model currently in use.

Step 4: Switch model via script (required)

Run:

bash <skill_dir>/scripts/switch-model.sh <hudl-model-id>

switch-model.sh is the source of truth and must be used instead of manual edits.

Expected behavior from script:

• Updates agents.defaults.model.primary
• Updates active agent model (agents.list[*].model.primary, preferring id: "main", otherwise first agent)
• Normalizes models.providers.hudl.models[*].id so every catalog ID is hudl/...
• Ensures the target model exists in models.providers.hudl.models
• Preserves unrelated configuration
• Fails if the resulting active/default model fields are empty, not hudl/-prefixed, or do not exactly match the requested target
• Fails if the provider model catalog still contains non-hudl/ IDs or is missing the target model

Failure handling:

• If script fails, stop and show the script error.
• Do not "repair" the config manually inside the skill. If the script does not leave config in a valid final state, stop and surface the failure.

Step 5: Verify post-switch values before restart (required)

Confirm both fields now point to the normalized target model and both are hudl/....

• If aligned: continue.
• If not aligned: report mismatch and stop. Do not restart.

Step 6: Restart + report final state (required)

Before restart, tell the user:

• restart is starting now
• service may take a couple of minutes to come back
• you will confirm applied model values after restart

Run:

openclaw restart

After restart, report:

• restart completion status
• previous active/default model values
• new active/default model values
• alignment status (aligned or mismatch)
• exact final model ID in use
• provider catalog normalization status

On "what model am I on" / "current model"

1. Run validation (scripts/validate.sh).
2. Read the resolved config (config.json or openclaw.json).
3. Report:

- current active agent model (agents.list[*].model.primary, prefer id: "main") - current default model (agents.defaults.model.primary) - whether they are aligned - whether each value is correctly hudl/-prefixed - config path used for the check

Reporting rules:

• Do not say "current model is X" based only on agents.defaults.model.primary.
• If agents.list[*].model.primary and agents.defaults.model.primary differ, say the active agent model is currently in use and the default is stale/misaligned.
• When the user asks to "show on file" or "prove it from config", show both paths and values:

- agents.list[main-or-first].model.primary - agents.defaults.model.primary

No restart is needed for read-only status checks.

Example exchanges

Switch to Opus:

User: switch to opus
Agent: Checking GRU gateway config... verified.
Agent: Resolved target model: hudl/claude-opus-4.6.
Agent: Current models -> active: hudl/minimax-m2.5, default: hudl/minimax-m2.5.
Agent: Running switch script... success.
Agent: Provider model catalog normalized and includes hudl/claude-opus-4.6.
Agent: Restarting gateway now; it may take a couple minutes.
Agent: Restart complete. Active: hudl/claude-opus-4.6, default: hudl/claude-opus-4.6 (aligned).

Mismatched config on status check:

User: what model am I on?
Agent: Config path: /home/ubuntu/.openclaw/openclaw.json
Agent: Active model in use: hudl/minimax-m2.5 (from agents.list[id=main].model.primary)
Agent: Default model: gpt-4.1 (from agents.defaults.model.primary)
Agent: Status: mismatch. The gateway is currently using hudl/minimax-m2.5; the default value is stale and should not be reported as the live model.

No hudl provider:

User: switch to opus
Agent: (validate.sh failed)
Agent: This skill requires the Huddle01 GRU gateway (hudl provider with baseUrl gru.huddle01.io) in your OpenClaw config. No changes were made.

Unknown model:

User: switch to deepseek-r2
Agent: Heads up: deepseek-r2 is not in the known catalog. I will set hudl/deepseek-r2; requests will fail unless that model exists on your GRU gateway.
Agent: Switching to hudl/deepseek-r2...