I open-sourced Codex Spark , a Codex plugin for delegating concrete Computer Use and Browser Use tasks to GPT-5.3 Codex Spark subagents. Repo: https://github.com/KingGyuSuh/awesome-codex-spark The Problem As Codex sessions get better at long-horizon reasoning, the bottleneck is not always "can the model click the button?" Often the better question is: Should the most reasoning-heavy model spend its context and tokens on mechanical UI work? If the parent session is doing architecture, code review, release verification, or product reasoning, I want it focused there. A visible-world task like opening a page, reading UI state, pasting approved content, or filling one approved form should be delegated as bounded execution. The Pattern Codex Spark uses one skill: $codex-spark-delegate .…