mlhher/late-cli

Stop degrading your model's reasoning. A minimal, zero-config AI coding agent. Enforced ephemeral subagents keep context pure. From tiny local models up to Sol, Fable and Kimi K3.

What it solves

Late addresses the "architectural bottleneck" of standard AI coding agents, where a single shared context window becomes cluttered with compile errors, lint failures, and file write logs. This "noise" fills the KV cache and degrades the model's reasoning capabilities over time.

How it works

Late uses a split-brain architecture that separates planning from execution:

  1. Main Orchestrator: A high-level agent that analyzes intent and forms a master plan. Its context window is kept pure, containing only instructions and definitive results.
  2. Ephemeral Subagents: The orchestrator spawns specialized subagents (e.g., for coding or research) to perform specific tasks. These subagents absorb the "noise" of execution (like raw data bloat or retry loops).
  3. Context Destruction: Once a subagent completes its task, its entire context is destroyed, and only the final outcome is returned to the orchestrator.

Who it’s for

It is designed for developers and builders who want high-throughput AI coding assistance with minimal setup, supporting both local LLMs (via llama-server) and cloud providers.

Highlights

  • Zero-Dependency Binary: Distributed as a statically compiled binary with no need for Python or Node.js environments.
  • Hybrid Model Routing: Ability to use a powerful reasoning model for planning and cheaper, faster models for execution.
  • Exact-Match Diffs: Uses strict search/replace blocks with self-healing to prevent silent file corruption.
  • MCP Integration: Natively supports Model Context Protocol servers via standard I/O.
  • Stateful Resilience: Maintains session history on disk to allow resuming work after reboots.
  • Git Worktree Support: Enables parallel agent instances across different branches without context bleeding.