lash

Terminal AI coding agent written from scratch in Rust. Works with any OpenAI-compatible API, OpenAI Codex Subscription, or Google Gemini.

Patch-based editing, shell execution, file search, web search, planning, skills, host-backed subagents, session resume/retry, and live token accounting.

What's different

• Two execution modes
  • rlm (default) — runs a persistent lashlang DSL runtime that keeps state across turns. The model emits lashlang programs that are evaluated locally, with parallel { } blocks for concurrent tool execution. This gives the agent a programmable scratch space and control flow beyond what the provider protocol offers.
  • standard — uses the provider's native tool-calling protocol directly. The model can emit multiple independent tool calls in a single response, which the runtime executes concurrently. Simpler, no extra DSL layer, closer to how the provider intended the API to be used.
• Plugin architecture — tools, prompts, planning, subagents, and history are all plugins; embedders compose what they need
• Embeddable — split into three crates so you can use only what you need:
  • lash-sansio — sans-IO turn machine. Pure logic: takes messages in, yields effects out, no networking or filesystem. Embed this if you want full control over I/O and just need the turn loop.
  • lash — async host runtime built on lash-sansio. Adds plugin infrastructure, tool execution, session management, context approaches, and child-session orchestration. Use this to embed a full agent in a larger Rust application.
  • lash-cli — the end-user binary. TUI, provider auth, config, and the single-shot --print mode. Built on top of lash.

Agent behavior

• Finish the task by continuing with the next concrete action while actionable work remains.
• Do not stop just to report that work is incomplete.
• Only summarize remaining work when blocked, when a decision is required, or when no further feasible action remains in the current turn.
• Do not claim completion unless the required end state has actually been verified.

Install

Download the relevant release, use the install script below or build from source.

curl -fsSL https://github.com/SamGalanakis/lash/releases/latest/download/install_lash.sh | bash

cargo build -p lash-cli --release

Quick start

lash                           # interactive TUI
lash -p "summarize this repo"  # single-shot, output to stdout

Agent behavior

• Continue with the next concrete action while actionable work remains.
• Do not stop merely to report that work is incomplete.
• Only summarize remaining work when blocked, when a decision is required, or when no further feasible action remains in the current turn.
• Do not claim completion unless the required end state has actually been verified.

Docs

See docs/index.html for architecture, design, and full reference.

License

MIT