The Agent Skills Directory

Persistence Mechanism (MEDIUM): The install-hook script (referenced in DEPLOYMENT.md) installs a permanent hook at ~/.claude/hooks/sessionEnd. This script automatically triggers the indexing process after every session. While this is the core functionality of the 'Remembering Conversations' skill, creating persistent execution hooks is a medium-risk behavior.
Sensitive Data Access (MEDIUM): The indexing system reads and processes historical conversation logs stored in ~/.config/superpowers/conversation-archive/. Accessing full conversation history is high-sensitivity; while necessary for the stated search functionality, it constitutes a significant data exposure risk if the agent is compromised.
Indirect Prompt Injection (LOW): The summarizer.ts logic and search-agent.md templates ingest untrusted data from past conversations. There is a risk that malicious instructions embedded in historical logs could influence the LLM during summarization or search synthesis.
Ingestion points: ~/.config/superpowers/conversation-archive/**/*.jsonl processed in parser.ts and summarizer.ts.
Boundary markers: Uses <summary> tags in summarizer.ts and structured markdown sections in search-agent.md.
Capability inventory: Local file read/write (fs), network access via the Anthropic SDK, and local command execution via shell scripts.
Sanitization: No explicit escaping or filtering of historical conversation content is performed before interpolation into prompts.
Command Execution (LOW): The sharing-skills component facilitates execution of git and gh (GitHub CLI) commands for branch management and PR creation, which involves remote interaction and credential usage.

meeting-insights-analyzer