The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to inspect system state, manage services, and manipulate files.
Evidence: Commands such as sqlite3, find, launchctl, docker exec, and chmod appear throughout SKILL.md and the references/ directory to manage the Hermes environment.
[DATA_EXPOSURE]: The skill directs the agent to access sensitive internal configuration and session data files.
Evidence: The instructions guide the agent to read ~/.hermes/config.yaml and ~/.hermes/state.db to extract memory entries, user profiles, and session history for analysis.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted session history from a database to drive system configuration changes (disabling skills).
Ingestion points: Reads the messages table from ~/.hermes/state.db (file: skill-pruning-workflow.md).
Boundary markers: None mentioned for processing the SQL query results.
Capability inventory: Shell execution (python3), file writing (config.yaml), and service management (launchctl).
Sanitization: Employs basic shell filtering (sed, sort -u) to normalize skill names before modification.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic script execution and describes runtime modification of agent internals through shims.
Evidence: Multiple Python one-liners (python3 -c) are used for YAML manipulation. Additionally, references/tool-router-architecture.md describes a plugin shim that wraps and replaces internal methods like agent.turn_context.build_turn_context at runtime.

hermes-context-optimization