Run long-horizon coding work with Late CLI
Use Late CLI to keep a lead coding agent focused while isolated researcher and coder subagents spend disposable compute on large implementation tasks.
npx skills add agentskillexchange/skills --skill run-long-horizon-coding-work-with-late-cli
Use Late CLI when an agent or operator needs to run a large coding task that would normally drown the primary agent in file reads, failed edits, compiler traces, and retry noise. The workflow is concrete: launch Late in a project, let the lead orchestrator plan and verify, delegate bounded research and implementation work to isolated subagents, optionally use a sandboxed worker path such as late-podman where supported, and review structured results and persisted subagent history before accepting changes.
What this skill actually does
Invoke this instead of using a normal single-agent coding loop when the job is long-horizon implementation, local-model coding, overnight refactoring, or multi-phase codebase repair where central context quality matters more than minimizing total worker tokens. It is especially useful when the operator wants enforced separation between planning and mutation: the orchestrator plans and verifies while workers execute inside fresh contexts.
This is not a generic local-LLM or CLI listing. The scope boundary is Late’s orchestrator/worker coding execution workflow, context preservation, subagent handoff, and optional isolated worker execution. Do not stretch it into broad model serving, generic agent frameworks, or ordinary terminal automation.
Inputs and prerequisites: Late CLI; a target Git repository; a supported cloud provider API key or local llama.cpp/llama-server setup; optional Podman/devcontainer runtime for isolated worker execution.
Setup notes: Install from the Homebrew tap with brew tap mlhher/late followed by brew install late on supported systems. From a project directory, run late; use late-podman only when the task should execute in an isolated rootless container and the local runtime is configured.
Source and verification boundary: use https://github.com/mlhher/late-cli and its upstream quickstart/docs as the canonical source before running the workflow; keep commands, model configuration, sandbox choices, and generated code changes reviewable against that source.
Framework fit: publish this as a Multi-Framework workflow because Late is model/provider agnostic and documents local and cloud model support rather than binding invocation to one host agent.