The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the 'all-MiniLM-L6-v2' embedding model from Hugging Face using the @xenova/transformers library for local processing.
[COMMAND_EXECUTION]: Executes indexing and search operations via shell scripts and tsx. It installs a hook in ~/.claude/hooks/sessionEnd to automatically run the indexer when an agent session terminates.
[DATA_EXFILTRATION]: Transmits conversation data to the Anthropic API to generate concise summaries for the search index. This is a primary feature of the skill.
[PROMPT_INJECTION]: Processes past conversation logs (untrusted data) to create embeddings and summaries. While this presents an indirect prompt injection surface, the skill implements a synthesis-based subagent workflow to reduce the risk of repeating malicious instructions.
Ingestion points: parser.ts reads .jsonl files from the project directory.
Boundary markers: The subagent template in prompts/search-agent.md requires synthesis and forbids raw excerpts.
Capability inventory: Access to shell command execution and Anthropic API integration.
Sanitization: Input is parsed as structured JSONL, though the text content is not filtered for natural language instructions.

remembering-conversations