CK Plan

Coordinate fresh worker plan → review → revise passes. Deliver a repository-grounded plan without implementing it; later implementers or review-loop may use it as optional input.

Inputs

Use user values or defaults: goal = current conversation (ask if absent); planning and review harness = current; model = select by role; review parallelism = 1; critic: auto; max cycles = unlimited within the run budget; output = display; propose_improvements = false. critic: always requires a fresh critic each cycle.

Before launch, record a finite elapsed wall-time allowance, starting at 60 minutes and adjusted to scope. Fit finite pass limits, scouts, revisions, reviews, critics, retries, validation, and reporting inside it. Honor user hard caps; limit tokens only with observable usage. Before a self-estimated allowance expires, permit a recorded finite extension citing concrete progress and remaining work. Never extend user caps or reset no-progress limits. Choose routine budgets autonomously; exhaustion ends incomplete with gaps and next actions.

Always plan the root-cause design so extra layers are unnecessary (“unwrap the onion”). With propose_improvements: true, collect advisory reusable instruction changes; modify skills only when separately asked.

Clarify decisions

When user intent is needed, invoke user-decision before continuing. For explicitly or implicitly noninteractive work, choose the best long-term options. Use quick options only when requested for the entire task.

Worker selection

Use fresh native subagents for current when supported, otherwise invoke-harness; use invoke-harness for named external harnesses. Supply each worker goal, role, focus, decisions, context, artifact index, output destination, and limits. Require one pass without re-entering this skill, nesting a harness, or waiting on stdin.