self-improvement

Pass

Audited by Gen Agent Trust Hub on Apr 21, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill manages internal developer workflows and project memory. All file operations and script executions are strictly project-local and do not involve remote data exfiltration, credential harvesting, or unverified remote execution.\n- [PROMPT_INJECTION]: The skill introduces a surface for indirect prompt injection by logging user feedback and tool outputs into a local history for future agent reference. This surface is managed through the use of delimited XML tags in automated reminder scripts.\n
  • Ingestion points: User corrections, command error outputs via error-detector.sh, and historical learning files in the .learnings/ directory.\n
  • Boundary markers: Automated reminders are wrapped in XML-like tags such as <self-improvement-reminder> and <error-detected> to assist the agent in distinguishing logged data from instructions.\n
  • Capability inventory: Local file read/write operations (SKILL.md, .learnings/*.md) and execution of shell scripts (scripts/extract-skill.sh, scripts/activator.sh) via the agent's hook system.\n
  • Sanitization: The extract-skill.sh script validates input using a strict regular expression to prevent arbitrary command injection or directory traversal.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 21, 2026, 05:34 AM