What problem does it solve? Long-running, vague, or stalled coding goals in Codex and Claude Code lose direction across turns: plans get re-invented, verification is skipped, and agents stop early. Goal Prep compiles raw user intent into a durable local goal board (goal.md, state.yaml, notes/) with a goal oracle, one active task, and receipt-based proof so runs can pause, resume, and finish against observable evidence. ## Core Features & Use Cases - Intake Compiler: Translates vague, specific, existing-plan, recovery, or audit input into a structured goal intake with completion proof, goal oracle, likely misfire, and blind spots, asking guided diagnostic questions when clarity is missing. - Role-Tagged Board Seeding: Creates a state.yaml task board with Scout (read-only evidence mapping), Judge (read-only decisions), and Worker (bounded writes inside allowed_files with verify and stop_if) tasks, plus a final audit task. - Validation and Receipt Tooling: Bundled scripts check board validity (check-goal-state.mjs), apply receipts atomically with fail-closed revert (apply-receipt.mjs), enforce a stop gate (check-can-stop.mjs), and dispatch tasks to external harness CLIs with git-based scope verification (dispatch-task.mjs). - Use Case: A user says "make this repo's billing flow reliable." Goal Prep runs diagnostic intake, creates docs/goals/billing-reliability/ with a charter and seeded Scout/Judge/Worker board, starts the local live board, and prints the exact /goal or /goalbuddy command to begin execution. ## Quick Start Ask the agent to run goal-prep for your broad coding objective, answer its intake questions, then start execution with the printed /goal or /goalbuddy command.
