The Agent Skills Directory

[COMMAND_EXECUTION]: The script orchestrator.py utilizes subprocess.run to invoke the local openclaw agent CLI for processing evolution tasks via a sub-agent. This is a core part of the skill's self-improvement logic.
[COMMAND_EXECUTION]: The apply-proposal.py script executes local git commands (git init, add, commit) to maintain a versioned history of all applied evolutions within the workspace, ensuring transparency and reversibility.
[PROMPT_INJECTION]: The skill implements an 'Evolution' workflow that scans user dialogue history (scan-history.py) to extract feedback signals. This creates an indirect prompt injection surface where a malicious user could attempt to influence future agent instructions. However, the risk is thoroughly mitigated by the design: all proposed 'mutations' to instructions must be explicitly reviewed and accepted by the user before being written to the filesystem.
[SAFE]: The skill operates entirely within the local filesystem. Data (including feedback signals, SQLite database, and reports) is stored in the ~/.arkclaw/workspace/evolution-data/ directory. Analysis of the scripts confirmed that no external network calls or exfiltration mechanisms are present.

byted-ark-evolve