The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interpolates session configuration data and task descriptions directly into the prompts used to spawn sub-agents.
Ingestion points: Data is loaded from .agenthub/sessions/{session-id}/config.yaml and template files in references/agent-templates.md.
Boundary markers: The prompt template in SKILL.md does not use delimiters or instructions to prevent sub-agents from obeying malicious commands embedded within the {task} or {variables} fields.
Capability inventory: Sub-agents are granted filesystem access (git worktrees), testing capabilities, and the ability to write result summaries to a shared directory.
Sanitization: There is no evidence of validation or sanitization for the interpolated strings before they are sent to the sub-agents.
[COMMAND_EXECUTION]: The skill executes a local Python script scripts/session_manager.py to update the state of the session. While this uses a local path, it represents a point of command execution based on session identifiers.

spawn