The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation in references/engine-design.md describes a mechanism for generating shell commands using templates (runner) and binding arguments. These commands are intended to be executed by agent lifecycle hooks, such as those in Claude Code or GitHub Copilot CLI.
[PERSISTENCE]: The skill's primary objective is to manage and install persistent configuration fragments into agent settings (e.g., ~/.claude/settings.json, ~/.hermes/agents-registry.yaml). These fragments define hooks that trigger specific publisher code automatically during agent operations.
[INDIRECT_PROMPT_INJECTION]: The framework creates an attack surface for indirect injection if a malicious master configuration file is processed by the synchronization engine.
Ingestion points: Configuration data is ingested from a user-provided *.master.json and a *.mappings.lock.json file as described in SKILL.md and references/engine-design.md.
Boundary markers: The skill recommends using a _do_not_edit header in generated files to prevent manual drift, but it does not specify explicit boundary markers or instructions to isolate generated command strings from malicious input during template interpolation.
Capability inventory: The system is designed to perform file system writes to sensitive agent configuration paths and generate shell commands for automated execution.
Sanitization: The engine design focuses on idempotency and determinism but does not detail sanitization or escaping mechanisms for arguments interpolated into shell command templates.

ssot-fanout