mlhher / late-cli

What it does

• Universal Plugin System. Extend Late with custom slash commands, MCP servers, themes, and lifecycle hooks (onSessionStart, onMessageSend, onToolCall, onToolResult). Install via late plugin install directly. Subagents automatically inherit a
• Empirical Logit Biasing. Built on Chenlong Wang et al., Findings of EMNLP 2025, Late dynamically suppresses repetitive self-reflection tokens ("Wait...", "Hmm") in reasoning models via llama-server, reducing CoT token bloat by 27%–51% with
• Overnight Autopilot via Podman (late-podman). Run the agent in an isolated, rootless container sandbox with devcontainer support and yolo mode. Turn Late loose on large-scale refactors overnight without risking your host machine.
• Parallel Subagents. Run independent plan steps as concurrent Coder/Researcher subagents (--max-async-subagents 2). Workers keep the same tool pruning and context isolation, all results return to the orchestrator in a single turn, and its ca
• Interactive Model Switching & Hybrid Routing. Press /model to reconfigure orchestrator and worker models on the fly. Route planning to frontier reasoning models while delegating execution to fast, cost-efficient workers.