The Agent Skills Directory

[COMMAND_EXECUTION]: The init-os.sh script and Python scripts (eval.py, learn.py) execute shell commands using git rev-parse to identify the project root. This is a standard development practice for workspace configuration.
[INDIRECT_PROMPT_INJECTION]: The skill implements a feedback loop where agents are explicitly directed to read .agent/learnings/*.json files and apply rule_change entries to their subsequent behavior. This creates a surface where instructions stored in data files can override agent behavior.
Ingestion points: .agent/learnings/, .agent/state/tasks.json, and .agent/evals/ (SKILL.md).
Boundary markers: None. Agents are told to "explicitly apply its rule_change entries to your approach."
Capability inventory: The skill provides Python scripts that write to the local filesystem and shell scripts that modify GEMINI.md and CLAUDE.md.
Sanitization: None. The content of rule_change is treated as a directive.
[COMMAND_EXECUTION]: The init-os.sh script performs file writes and appends instructions to GEMINI.md and CLAUDE.md. It also attempts to automate the execution of a heartbeat skill by injecting shell commands (python3 ~/.claude/skills/heartbeat/scripts/heartbeat.py pop) into these project documentation files.

agentic-os