v0.1.1

Fixed

• BEAMCLAW_USER bypassing Telegram pairing access control

Changed

• Route Nano Banana Pro image generation through OpenRouter
• Add retry logic for Gemini text-only responses in generate_image.py
• Strip old image attachments from LLM context to reduce token usage
• Strengthen skill prompt to prevent unnecessary file verification

Added

• Save incoming Telegram photo attachments to disk for tool access

BeamClaw v0.1.0

Changelog

[0.1.0] — 2026-03-05

First public release.

Core

• Agentic loop (gen_statem) with streaming, tool execution, approval workflow
• LLM providers: OpenRouter (default), OpenAI
• Session persistence via Mnesia (survives restarts and Docker rebuilds)
• Context compaction with pre-compaction memory flush
• Credential scrubbing on all tool results
• Token-based automatic compaction trigger

Tools

• Built-in: bash, terminal, curl, jq, read_file, write_file, delete_file
• Workspace memory (MEMORY.md, daily logs, bootstrap files, search)
• Web search (Brave Search API)
• Sandboxed code execution (Docker, with MCP tool bridge)
• Scheduled tasks and heartbeat

Channels

• Telegram (long-poll + webhook, photo/vision, voice transcription, pairing)
• TUI (local and remote via Erlang distribution)
• HTTP API (OpenAI-compatible, SSE streaming)
• WebSocket

Agent System

• Named agent workspaces with six bootstrap files
• Skill system (bundled + global + per-agent, BM25 auto-injection)
• Per-user agent mapping via Telegram pairing

Search

• BM25 keyword search (pure Erlang)
• Vector semantic search (OpenAI-compatible embeddings)
• Hybrid score merging

Infrastructure

• Nine OTP apps in rebar3 umbrella
• Docker image (Alpine, multi-stage, <100 MB)
• CLI escript (30 commands)
• 720 automated tests (683 EUnit + 37 CT)