coach

Pass

Audited by Gen Agent Trust Hub on Apr 27, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill defines a feedback loop that ingests untrusted data from tool execution outputs and user corrections to generate persistent agent instructions. 1. Ingestion points: Friction signals are captured from stderr patterns, user correction language, and session context into a local SQLite database. 2. Boundary markers: The system prompt does not define specific delimiters or instructions to ignore nested commands when processing these signals. 3. Capability inventory: The skill uses Bash with python3 and sqlite3 to process data and has Write permissions for CLAUDE.md to persist rules. 4. Sanitization: No explicit sanitization or filtering of captured signals is described. This risk is mitigated by the requirement for manual user approval before rules are applied.
  • [COMMAND_EXECUTION]: The skill relies on executing a suite of local Python scripts (init_coach.py, detect_signals.py, etc.) and SQLite queries to manage the learning lifecycle. These tools are used for internal data aggregation and signal analysis within the skill's environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 27, 2026, 12:46 AM