traul

CLI tool that watches communication streams (Slack, Telegram, Discord, Linear, Gmail, Claude Code sessions, Markdown files, WhatsApp), indexes messages, detects patterns via signals, and surfaces actionable insights. Runtime: Bun + TypeScript | DB: SQLite (WAL mode, FTS5, sqlite-vec) | Embeddings: node-llama-cpp (Qwen3-Embedding-0.6B), Ollama fallback | Version: 0.2.0 Project: /Users/dandaka/projects/traul ---

Sync messages from communication sources incrementally. | Argument | Description | |----------|-------------| | source | Optional. slack, telegram, discord, linear, gmail, claudecode, markdown, or whatsapp. Omit to sync all. | • Uses cursor-based incremental sync (only fetches new messages) • Slack: syncs channels + thread replies, caches user profiles, stores reactions/reply_count in metadata • Telegram: syncs via tg.py (Telethon), 1-hour cooldown skip per chat, progress reporting every 10s • Claude Code: reads all JSONL sessions from ~/.claude/projects/, extracts user+assistant messages • Markdown: walks configured dirs for .md files, tracks changes by content hash • Reports messages added + contacts discovered per source

Hybrid search combining vector similarity (semantic) and FTS5 keyword matching with Reciprocal Rank Fusion. Falls back to FTS-only if embedding is unavailable. Search modes: • Hybrid (default) — best for multi-word and exploratory queries. Finds semantically related messages even when exact keywords don't appear. Uses node-llama-cpp with Qwen3-Embedding-0.6B (auto-downloads ~639MB model on first use). Falls back to Ollama, then FTS-only. Prints coverage ratio to stderr (e.g. 88% vector, 12% FTS). • FTS-only (--fts) — keyword matching with BM25 ranking. Faster, but requires ALL terms to match (implicit AND). Brittle with multi-word queries, especially combined with source/channel filters. • OR mode (--or) — joins search terms with OR instead of AND. Works with both --fts and hybrid. Use for broad exploratory queries where any term is relevant. • Substring (--like) — bypasses FTS entirely, uses SQL LIKE. Useful for exact phrases that FTS tokenization breaks (e.g. "how do I"). Tip: Prefer hybrid (default) for broad queries like "metrics mixpanel registration". Use --fts --or for exploratory keyword searches matching ANY term. Use --like for exact phrase matching. | Option | Description | |--------|-------------| | -s, --source <source> | Filter by source | | -c, --channel <channel> | Filter by channel name | | -a, --after <date> | Messages after ISO date (aliases: --from, --start) | | -b, --before <date> | Messages before ISO date (aliases: --to, --end) | | -l, --limit <n> | Max results (default: 20) | | --json | Output as JSON | | --fts | Keyword-only search (skip vector search) | | --or | Join terms with OR instead of AND (works with --fts and hybrid) | | --like | Substring match (LIKE) — bypasses FTS, useful for exact phrases |

Retrieve a full conversation thread by its thread ID, or list all threads from a given date. Search results display thread IDs in the output so you can copy them for use with get. | Option | Description | |--------|-------------| | thread-id (positional) | Thread ID (e.g. Claude Code session UUID) | | -d, --date <date> | Get all threads from a date (ISO 8601) | | --json | Output as JSON | `bash