supervise-plan

Warn

Audited by Gen Agent Trust Hub on Jul 7, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Employs AppleScript (osascript) to automate the Ghostty terminal emulator. This allows the skill to programmatically open new terminal tabs and inject shell commands, such as claude "/execute-plan ...", into the user's active shell environment.
  • [DATA_EXFILTRATION]: Accesses sensitive session transcripts stored in ~/.claude/projects/. The skill reads JSONL files containing the full history of interactions between the user and the agent to calculate context window usage.
  • [COMMAND_EXECUTION]: Executes dynamic Python code via the shell (python3 -c) to parse internal application data and environment variables such as CLAUDE_CODE_SESSION_ID.
  • [REMOTE_CODE_EXECUTION]: Utilizes an autonomous loop mechanism to persistently reschedule its own execution every 10 minutes, effectively creating a background daemon without direct user oversight for each wake cycle.
  • [PROMPT_INJECTION]: Ingests status lines and headings from local files (plan.md, mailbox.md) to drive its control logic. This creates an attack surface for indirect prompt injection if those files are modified by untrusted third-party processes or external data.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 7, 2026, 03:12 PM
Security Audit — agent-trust-hub — supervise-plan