coder-leo changelog

#2.0.0 - 2026-10-20

• First public release: a coding agent for the terminal on local models, through Ollama or any OpenAI-compatible server (LM Studio, llama-server, vLLM, mlx)
• Tools named as in Claude Code: Read, Write, Edit, LS, Glob, Grep, Bash (with background jobs), WebFetch, WebSearch, TodoWrite, Task, AskUserQuestion, ExitPlanMode, plus Search, Outline, RepoMap, Recall and Skill; the names other agents use are understood too
• Context window kept in budget: automatic size from the model and the memory, calibrated counting, tool results cut to what fits with the full output saved to a file, old results cleared at 70%, a structured summary at 85% (in parts when the conversation is too long), the transcript searchable with Recall, /context to see what uses the window
• Repository map in the system prompt and the Search tool for plain-language questions about the code
• Permission modes default, acceptEdits, plan and yolo, rules in the settings, destructive commands guarded; Shift+Tab cycles the mode
• Sessions saved as they happen, continued with -c, resumed by id or from a list, exported as Markdown; file checkpoints and /rewind
• CLAUDE.md and AGENTS.md instructions with imports, custom commands, skills, sub-agents, hooks and MCP servers, read from .coder-leo and from .claude directories
• Streaming Markdown with syntax highlighting, diffs before edits, a task list, a status line, a welcome box, history search, big pastes folded, images for vision models, thinking levels
• Headless mode with text, json and stream-json output
• Installs with Homebrew on macOS and with one command on Linux, Android (Termux) and WSL; coder-leo update brings new versions