The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local, bundled Python script to identify agent environments on the system.
Evidence: SKILL.md instructs the agent to run python3 -I scripts/detect-harnesses.py to discover candidate harnesses.
[PERSISTENCE]: The primary function of the skill is to perform long-term modification of the agent's behavior by injecting instructions into global steering files.
Evidence: SKILL.md describes appending or replacing a delegation policy in files like AGENTS.md, CLAUDE.md, and GEMINI.md, which are read by agents on every turn.
[DATA_EXFILTRATION]: The skill scans the user's home directory to map out AI toolchain configurations.
Evidence: scripts/detect-harnesses.py searches Path.home() and ~/.config for directories containing specific agent-related files (e.g., mcp.json, steering files).
[INDIRECT_PROMPT_INJECTION]: The skill processes existing configuration files and external harness documentation which could theoretically contain malicious instructions.
Ingestion points: scripts/detect-harnesses.py reads the content of existing steering files to check for the presence of the delegation policy; the agent is instructed to read harness-specific documentation to determine schemas.
Boundary markers: The skill uses a specific Markdown heading ## Delegation and model routing as a delimiter to manage the injected section.
Capability inventory: The skill can execute local scripts, read and write files within the user's home directory, and create new subagent files.
Sanitization: The injected content is sourced from a static file (POLICY.md), reducing the risk of dynamic injection, although the agent interprets the output of the detection script and external documentation during the process.

setup-delegation-policy