The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing untrusted 'Evolution goal' inputs and incorporating them into instructions for subsequent agent tasks.
Ingestion points: User-provided Evolution goal and Target inputs defined in SKILL.md.
Boundary markers: The templates for brainstorming (Phase 1) and delegation (Phase 3) do not utilize explicit boundary markers or delimiters to isolate user-supplied goals from the agent's instructions.
Capability inventory: The skill has access to Bash, Read, and Write tools, and it delegates tasks to the copilot-cli-agent with instructions to write files directly to the filesystem.
Sanitization: There is no evidence of input validation, sanitization, or filtering of the user-provided goals before they are interpolated into the task plans.
[COMMAND_EXECUTION]: The skill uses shell commands and local scripts to orchestrate its workflow and manage logging.
Evidence: Execution of run_agent.py to invoke different model backends for brainstorming and dispatching tasks.
Evidence: Use of standard utilities such as grep, wc, and test to perform heartbeat checks and verify the integrity of dispatch outputs.
Evidence: Execution of plugins/agent-agentic-os/scripts/experiment_log.py to record evolution metadata.

os-evolution-planner