mlhher/late-cli

What it solves

Late is an AI coding agent designed to prevent the "reasoning collapse" that occurs when an LLM's context window becomes cluttered with execution noise—such as compiler errors, file reads, and failed diffs. By isolating the planning and execution phases, it allows models to handle tasks that exceed their native context limits without losing accuracy.

How it works

Late uses an architectural split between a Lead Orchestrator and ephemeral subagents:

• Lead Orchestrator: Strictly handles planning and verification. It is physically prevented from writing files or executing mutating bash commands.
• Subagents (Coder and Researcher): Spawned into isolated contexts to perform atomic tasks. Once a task is finished, the subagent's noisy scratchpad is wiped, and only high-signal structured diagnostics are returned to the orchestrator.
• Logit Biasing: For llama.cpp users, it suppresses redundant "thinking" tokens (e.g., "Wait...", "Hmm") to reduce Chain-of-Thought bloat.
• Tool Pruning: Hard boundaries are enforced so the orchestrator cannot bypass delegation, and subagents cannot spawn further agents.

Who it’s for

Developers who use local LLMs (via llama-server) or cloud providers to perform complex, multi-step coding tasks, particularly those working with large codebases where monolithic agents typically fail due to context pollution.

Highlights

• Architectural Isolation: Enforces a strict separation between planning and execution to preserve the orchestrator's context.
• Zero-Dependency Binary: Written in Go for near-instant startup (<10ms) without requiring Python or Node.js runtimes.
• Sandboxed Execution: Supports rootless devcontainers via late-podman for safe, autonomous overnight refactors.
• Model Agnostic: Works out-of-the-box with llama-server (port 8080) or major cloud APIs (DeepSeek, Claude, GPT, etc.).
• Extensible: Includes a universal plugin system for custom slash commands, MCP servers, and lifecycle hooks.