v0.1.0

Local RAG search over Claude Code session history. Verbatim (no LLM
summarization), nothing injected into sessions, secrets redacted before storage.

Added

• Hybrid search: ONNX bge-m3 semantic embedding + FTS5 trigram keyword search fused with RRF
• MCP server with search, expand, read_session, and list_sessions tools
• Claude Code Stop/SessionEnd hooks that only index sessions silently — nothing is summarized or injected back into any session
• Verbatim recall: indexes cleaned transcript turns (no LLM summarization step); read_session/expand serve the stored, redacted turns so recall works even after the original transcript is pruned or moved
• Secret redaction before storage across all content (user/assistant text, tool-call summaries, chunk text and titles): API keys/tokens (incl. Anthropic sk-ant-), compound secret keys (client_secret, aws_secret_access_key, …), Bearer tokens, PEM private keys, JWT, env-var assignments, home-path normalization. Redaction is mandatory (not configurable)
• redact-db command: re-applies redaction to all stored turns and chunks (incl. titles and the FTS shadow) in place, idempotently
• Project filtering for scoped searches; migrate command for project path renames
• Incremental indexing keyed on session id, SHA-256, mtime, and an index signature (embedding model, dimension, chunk size, redaction version) so changing any of them re-indexes affected sessions instead of skipping them
• Atomic session indexing (single transaction, rollback on failure); the database directory and file are created owner-only (0700/0600); MCP server loads the embedding model once per process
• Configurable embedding model, database path, and chunking via TOML