Understanding the `disable-model-invocation` Flag in the i-have-adhd Repository

The disable-model-invocation flag is a YAML directive in SKILL.md that prevents automatic activation of skill rules, requiring explicit user invocation (such as /i-have-adhd) before the runtime applies any behavioral modifications.

The ayghri/i-have-adhd repository implements a skill system designed to shape AI outputs for readers with ADHD. Within this architecture, the disable-model-invocation flag serves as a critical control mechanism that determines when skill behaviors become active across different AI runtimes.

How the disable-model-invocation Flag Works

The flag resides in the YAML front-matter of the skill definition. In skills/i-have-adhd/SKILL.md at line 4, setting disable-model-invocation: true tells compatible runtimes to load the skill description while withholding the actual behavioral rules until explicitly requested.

Skill Definition Configuration

When authoring a skill, you declare the flag in the metadata header:

# skills/i-have-adhd/SKILL.md

---
name: i-have-adhd
description: 'Shape output for a reader with ADHD…'
disable-model-invocation: true   # Prevents automatic activation

license: MIT
---

This single line instructs the runtime harness to treat the skill as inactive by default, even after installation.

Runtime-Specific Handling

Different AI runtimes interpret this directive through distinct mechanisms:

• Claude Code and Qwen Code: These environments read the flag directly from SKILL.md and maintain a silent state until the user issues the explicit invocation command.
• Codex: This runtime ignores the SKILL.md flag entirely. Instead, it respects the allow_implicit_invocation: false policy defined in skills/i-have-adhd/agents/openai.yaml at lines 6-7.
• Third-party plugins: Some custom runtime hooks, such as those implemented in hooks/always-on.mjs, may load skill definitions at startup and activate them automatically regardless of the flag status.

Why Some Runtimes Stay Silent Until Invoked

When the disable-model-invocation flag (or its Codex equivalent) is active, the runtime performs a "silent installation." As documented in INSTALL.md at line 805, the system loads the skill's description into memory but does not inject any rules into the model's prompt context. The model behaves as if the skill does not exist, producing standard responses without ADHD-specific formatting or constraints.

This silence persists until the user explicitly invokes the skill (e.g., typing /i-have-adhd), at which point the runtime injects the skill's rules into the active context. This behavior was verified in pull request #60, where testing confirmed that Copilot (a Codex-based runtime) applied no skill rules until the explicit command was issued.

Configuration Examples by Runtime

Standard Skill Definition

For Claude Code and Qwen Code compatibility, set the flag in the skill metadata:

# skills/i-have-adhd/SKILL.md

---
name: i-have-adhd
description: 'Action-first output for ADHD readers'
disable-model-invocation: true
---

Codex Agent Configuration

For OpenAI Codex, configure the equivalent policy in the agent definition:

# skills/i-have-adhd/agents/openai.yaml

interface:
  display_name: "I Have ADHD"
  short_description: "Action-first output for ADHD readers"
policy:
  allow_implicit_invocation: false   # Codex uses this instead

Explicit Invocation Flow

Before invocation:

User: How do I improve my code style?
Assistant: [Standard response without ADHD-specific formatting]

After invocation:

User: /i-have-adhd
Assistant: 1. Open `src/main.ts`
           2. Add the configuration line as shown above
           Next: run `npm test` to verify

Summary

• The disable-model-invocation flag in SKILL.md controls whether a skill activates automatically or waits for explicit invocation.
• Claude Code and Qwen Code read this flag directly to determine activation behavior.
• Codex ignores the flag and uses policy.allow_implicit_invocation in agents/openai.yaml instead.
• When active, the runtime loads the skill description but remains silent, applying no rules until the user explicitly invokes the skill (e.g., /i-have-adhd).
• This architecture gives skill authors precise control over when behavioral modifications affect model output.

Frequently Asked Questions

What is the purpose of the disable-model-invocation flag?

The flag gives skill authors control over activation timing. By setting disable-model-invocation: true in SKILL.md, authors prevent the runtime from automatically applying skill rules to every model response, ensuring the skill only activates when the user explicitly requests it.

Which runtimes respect the disable-model-invocation flag?

Claude Code and Qwen Code read the flag directly from SKILL.md. Codex ignores the flag and instead checks allow_implicit_invocation in agents/openai.yaml. Third-party runtime implementations may vary, with some custom hooks loading skills automatically regardless of the flag setting.

Why does my runtime stay silent after installing the skill?

The runtime remains silent because the flag (or equivalent policy) is active. According to the installation documentation at line 805 of INSTALL.md, the system loads the skill description but withholds all behavioral rules until you explicitly invoke the skill using the designated command (such as /i-have-adhd).

How does Codex handle implicit invocation differently than other runtimes?

Unlike Claude Code and Qwen Code, which parse disable-model-invocation from SKILL.md, Codex references policy.allow_implicit_invocation in skills/i-have-adhd/agents/openai.yaml. When set to false at lines 6-7, Codex maintains the same silent behavior but through a separate configuration mechanism.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}