tolibear/goalbuddy

What it solves

GoalBuddy provides a structured operating loop for long-running coding tasks that are too large for a single prompt. It prevents AI agents (specifically Codex and Claude Code) from losing orientation or re-inventing plans during complex projects, such as broad project improvements, bug hunts, or large refactors.

How it works

GoalBuddy creates a local workspace within the repository consisting of a charter, a goal oracle (a verification signal like a test suite), a board, and notes. It uses a state-tracking system (state.yaml) to maintain the source of truth across turns.

The workflow follows a specific logic: Intent $\to$ Oracle $\to$ Surface $\to$ Loop $\to$ Proof. It employs specialized roles—Scout (mapping the repo), Judge (slicing work into safe, useful chunks), and Worker (completing tasks and leaving receipts)—to ensure work is bounded and verified.

Crucially, it is harness-neutral, meaning a goal can be started in one AI tool (e.g., Codex) and resumed in another (e.g., Claude Code) because the state is stored as plain files in the repo.

Who it’s for

Developers using AI coding agents like Codex and Claude Code who are managing complex, multi-step software engineering tasks that require persistence and verification.

Highlights

• Cross-Harness Compatibility: Start a goal in one agent and resume it in another without losing progress.
• Local Live Boards: A visual interface to track plans, active tasks, and receipts without digging through chat history.
• Oracle-Based Verification: Requires a concrete signal (like a test suite) to define success, preventing aimless agent loops.
• Safe Slicing: Optimizes for the largest possible safe and reversible work slice rather than tiny, trivial tasks.