Description:
According to the documentation:
│ "Skills are progressively loaded — only metadata (name and description) is loaded at startup, with full content loaded on demand when the agent determines it's
needed."
Expected behavior: Agent sees skill descriptions and loads full content when user query matches.
Actual behavior: Agent never loads skill content automatically. It doesn't appear to see the available skill descriptions.
Evidence:
- Created a 6.7MB skill with unique, fictional content:
yaml
name: ccsm-zephyr-tc-execution
description: Guide for executing Zephyr test cases in CCSM. Use when user asks about running Zephyr tests, TC execution, test case CCSM-TC-7742, or the Zephyr integration process.
-
Skill contains invented "Golden Sequence" (INIT → PREFLIGHT → EXECUTE → HARVEST → SUBMIT) that the model cannot know.
-
Asked: "How do I execute test case CCSM-TC-7742?"
-
Agent gave generic response, never mentioned the Golden Sequence or any skill content.
-
Context stayed at 2.7% - skill was never loaded.
Environment:
- kiro-cli 1.24.1
- Skill file structure: .kiro/agents//skills//SKILL.md
- Frontmatter with name and description present
Conclusion: The skill:// scheme prevents initial loading (good), but the on-demand loading mechanism doesn't work - the agent has no way to see available skills or
request their content.
Description:
According to the documentation:
│ "Skills are progressively loaded — only metadata (name and description) is loaded at startup, with full content loaded on demand when the agent determines it's
needed."
Expected behavior: Agent sees skill descriptions and loads full content when user query matches.
Actual behavior: Agent never loads skill content automatically. It doesn't appear to see the available skill descriptions.
Evidence:
yaml
name: ccsm-zephyr-tc-execution
description: Guide for executing Zephyr test cases in CCSM. Use when user asks about running Zephyr tests, TC execution, test case CCSM-TC-7742, or the Zephyr integration process.
Skill contains invented "Golden Sequence" (INIT → PREFLIGHT → EXECUTE → HARVEST → SUBMIT) that the model cannot know.
Asked: "How do I execute test case CCSM-TC-7742?"
Agent gave generic response, never mentioned the Golden Sequence or any skill content.
Context stayed at 2.7% - skill was never loaded.
Environment:
Conclusion: The skill:// scheme prevents initial loading (good), but the on-demand loading mechanism doesn't work - the agent has no way to see available skills or
request their content.