First release since v1.5.0. Adds a self-learning loop, knowledge migration from other assistants, opt-in token saving, and replaces the custom Telegram bridge with the official Claude Code plugin.
✨ New
Self-learning — Opt in during setup and paw gets better over time. After each session a Stop hook runs a cheap heuristic gate (≥5 tool calls, a recovered error, or a correction you made); only then does it spend one small claude -p call to distill a single durable takeaway into a learnings file. Never stores secrets or full transcripts. Disable anytime with OPENPAW_SELF_LEARNING=off.
Selectable storage — Choose where learnings live in the wizard: plain markdown (~/.claude/memory/learnings.md, zero setup), an Obsidian vault, or a custom path.
openpaw migrate — Import what another assistant already knows. Detects Hermes, OpenClaw, GitHub Copilot, and Claude Code history, pulls their facts/persona into self-learning memory (secrets stripped on the way in), and can seed your SOUL.md. Preview with openpaw migrate --dry-run; offered automatically in the wizard when another assistant is detected.
Token saving (rtk) — Opt-in wizard step wires up rtk (Rust Token Killer) as a PreToolUse hook that compresses verbose command output for 40–90% fewer tokens.
♻️ Changed
Telegram — Replaced the homegrown bot bridge with the official Claude Code Telegram plugin (telegram@claude-plugins-official). OpenPaw installs + configures it (bot token + allowlist); the bridge runs inside Claude Code with the same model, skills, and permissions — no separate process to drift or crash.
🧹 Internal
- Dropped
grammy/@grammyjs/hydratedependencies. - README + docs updated.