Remembering Conversations

Pass

Audited by Gen Agent Trust Hub on Mar 13, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted data from historical conversation logs, which presents a surface for indirect prompt injection attacks.\n
  • Ingestion points: Reads conversation files in JSONL format from ~/.claude/projects/ and ~/.config/superpowers/conversation-archive/ (see indexer.ts and parser.ts).\n
  • Boundary markers: Employs specific system prompts in summarizer.ts to enforce a factual, non-conversational summary format (e.g., "Output ONLY the summary
  • no preamble").\n
  • Capability inventory: The tool chain allows reading project logs, writing to local archive directories, and making network requests to the Anthropic API via the Claude Agent SDK.\n
  • Sanitization: Uses hierarchical summarization in summarizer.ts to process large logs in chunks, which reduces the likelihood of an embedded injection successfully influencing the final synthesized output.\n- [SAFE]: Network activity is restricted to well-known and trusted services, including Anthropic's API for summarization and Hugging Face for model weights.\n- [SAFE]: The auto-indexing persistence mechanism (installed via install-hook) is documented and includes a safe installation process that creates backups and supports merging with existing hooks.\n- [SAFE]: No obfuscated code, hidden URLs, or hardcoded credentials were found during the analysis.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 13, 2026, 06:59 AM