The Agent Skills Directory

[PERSISTENCE]: The tool/install-hook script installs a persistent hook at ~/.claude/hooks/sessionEnd. This hook automatically triggers the indexing process in the background after every Claude Code session concludes.
[EXTERNAL_DOWNLOADS]: The skill uses the @xenova/transformers library to download the Xenova/all-MiniLM-L6-v2 embedding model from Hugging Face during its initial run to enable local vector search.
[DATA_EXFILTRATION]: Conversation logs from ~/.claude/projects (which may contain sensitive code, architectural decisions, or credentials) are sent to Anthropic's APIs to generate AI summaries that facilitate search.
[INDIRECT_PROMPT_INJECTION]: The skill indexes arbitrary conversation history which could contain malicious instructions. These instructions could attempt to influence the 'Search Agent' subagent's behavior when it processes and synthesizes search results.
Ingestion points: Reads .jsonl files from ~/.claude/projects/ and ~/.config/superpowers/conversation-archive/.
Boundary markers: The tool/prompts/search-agent.md template includes 'Critical Rules' instructing the agent to synthesize findings and avoid raw conversation excerpts to minimize context pollution.
Capability inventory: Background shell execution via npx tsx, file archival/writes, and network requests to Anthropic's API for summarization.
Sanitization: No explicit sanitization of conversation content before it is processed by the summarizer or indexed.
[COMMAND_EXECUTION]: The skill utilizes the sqlite-vec package, which involves loading a native binary extension into the SQLite database at runtime to perform vector similarity calculations.

Remembering Conversations