goal-loop

# Goal Loop A reusable, goal-driven improvement loop. The human runs `/goal-loop <target> <intent>` to name **one thing to improve** and what "better" means; the two agents (Claude + Codex) **plan and set the goal themselves**, break the target into testable stories, validate them with real execution, improve whatever the goal covers (correctness, behavior, performance, UX, polish), and ship autonomous PRs — looping until the goal's quality bar and the peer both pass. Design is one possible dimension, not the focus: route UI work through `impeccable` only when the goal touches the interface. The human never hand-authors a `/goal` prompt. The skill owns goal generation: it synthesizes a Scope Contract from the human's target/intent, then a **Goal Block** whose first token is literally `/goal ` when handed to Codex's goal loop. ## Roles - The agent the human invokes is the **coordinator-writer**. It owns the planning files and applies code changes. - The other agent is the **peer**: adversarial review + runtime validation. It writes code only under a recorded **lease** (see Leases). - **Both agents test in parallel**, partitioned by area, each on its own browser / computer-use session. - Subagents inherit the role of the agent that spawned them and the same skill-reading requirement. Default invocation is in the **Claude pane** (Claude = coordinator). The skill works either way; the only asymmetry is the Codex goal kickoff (below). ## Required reading (every agent...