The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell commands to manage its lifecycle and environment. This includes process management using pkill (targeting mempalace.mcp_server and mempalace mine), file operations via rm and cp for staging data, and search operations using grep and gh search. It also utilizes git for version control and the hermes CLI for task triggering.
[DYNAMIC_EXECUTION]: A core feature of this skill is the automated authoring of new SKILL.md files which are written to the agent's local registry at HERMES_SKILL_DIR. These files serve as executable instructions for the agent, meaning the skill dynamically generates its own functional extensions at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its data processing pipeline:
Ingestion points: It reads raw conversation history from ~/.hermes/sessions/ and mines it into a memory palace.
Boundary markers: The skill does not implement explicit boundary markers or sanitization when reading these logs or when citing 'drawers' as evidence for new skills.
Capability inventory: The skill has the ability to write new instruction files to the filesystem, execute shell commands, and interact with external GitHub repositories.
Sanitization: There is no evidence of sanitization for the content mined from sessions before it is used to template new skill instructions.
[EXTERNAL_DOWNLOADS]: The skill uses the GitHub CLI (gh) to search for existing skills in remote repositories. While interacting with a well-known service, this involves querying external data sources during its discovery phase.
[DATA_EXFILTRATION]: If configured, the skill can automatically create Pull Requests on GitHub (gh pr create) containing diary entries that summarize agent activity, including successes, failures, and identified capability gaps derived from private session logs.

electric-sheep