planning-with-files

Pass

Audited by Gen Agent Trust Hub on May 20, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill utilizes start and stop hooks to execute local shell and PowerShell scripts (check-complete.sh, check-complete.ps1) for task state verification. It also encourages the manual execution of a Node.js script (session-catchup.mjs) for context recovery.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its session recovery mechanism.
  • Ingestion points: The session-catchup.mjs script reads conversation history from .jsonl files located in the ~/.claude/projects/ directory.
  • Boundary markers: The script uses simple text headers such as --- UNSYNCED CONTEXT --- to delimit historical data, which may not prevent a model from following instructions embedded in the logs.
  • Capability inventory: The skill possesses extensive capabilities including Bash execution and file manipulation (Write, Edit, Read) across its various scripts and hooks.
  • Sanitization: The script performs basic JSON parsing and string slicing but lacks robust sanitization to filter out instruction-like content from the ingested session history.
Audit Metadata
Risk Level
SAFE
Analyzed
May 20, 2026, 05:04 AM
Security Audit — agent-trust-hub — planning-with-files