Manual-Only Skills: Locking Agent Actions to Humans
How the disable-model-invocation skill creates manual-only agent actions, with setup and practical scenarios.
2026-09-04
保留发布时的原文,未自动翻译。
Where this fits
When your agent has tools for destructive or privileged operations, the default "act first, ask later" behavior can be dangerous. You want to keep a skill visible and documented but make it impossible for the model to call it on its own. This skill does exactly that: it publishes a manual-only command that only a human can trigger, so it stays out of autonomous loops.
Why agents benefit
- The model can see the skill exists but is blocked from auto-invoking it, lowering the risk of unintended side effects.
- A human can still invoke it with
/skill:disable-model-invocation, so emergency actions don't require editing agent config mid-task. - It creates a clear permission boundary in the skill registry without building a separate approval system.
- It works as a reusable demonstration of the
disable-model-invocationflag, making it easy to copy into your own skills. - It lets agents surface the skill as a suggestion to the user instead of executing it, keeping the human in the loop.
Practical scenarios
Production database rollback
A developer asks the agent to investigate a bad migration. The agent suggests a rollback skill but cannot run it; the developer reviews the command and invokes /skill:disable-model-invocation only after confirming.
Sensitive customer-data export
While debugging a support issue, the agent could retrieve customer data but the skill is manual-only. The operator triggers it in a controlled session instead of letting the model pull data during a normal exploration.
Break-glass incident runbook
In an outage, an operator needs to restart infrastructure or revoke tokens. The manual-only skill remains available as a documented command, so the human can execute it quickly while the agent focuses on diagnosis.
Add it to your agent workflow
Install the skill with:
npx skills add Wholiver/metis --skill disable-model-invocation
After installation, the skill is registered but the model will not invoke it. To use it, the human enters:
/skill:disable-model-invocation
No special configuration is required beyond keeping the skill in your agent's skill directory.
Compare before adopting
Before making this part of every agent setup, compare it with:
- Quality signals: Check the repo's issue tracker and usage examples for how the flag is parsed by different agent runtimes.
- Maintenance freshness: The skill is small, so look at recent commits to see whether it still tracks changes in skill-loader behavior.
- Alternatives: Some runtimes support permission rules, read-only mode, or manual approval prompts; decide whether a separate manual-only skill is cleaner than those.
- Workflow fit: This is most useful for high-risk commands. If you need every action gated, a broader policy layer may be better.
Why it is worth tracking
The pattern is more important than the code: it shows how to separate what the agent knows about from what the agent can do. That distinction is essential for production agents. With 134 stars and a simple, focused README, it's a good signal that developers want explicit human-in-the-loop controls. It is worth evaluating when you design permission boundaries for irreversible actions. You can track it on OpenAgentSkill alongside other agent skills at https://www.openagentskill.com/skills/wholiver-disable-model-invocation.