This project addresses the challenge of autonomously planning and executing complex software development tasks end-to-end. It targets engineers and power users working with AI coding assistants like Claude Code and Codex CLI, offering a significant reduction in human interaction from numerous turns to a single approval and paste, thereby accelerating development cycles.

How It Works

Supergoal transforms task execution by first performing deep planning, including codebase recon, risk assessment, and memory integration. It then decomposes the task into an adaptive number of phases, generating detailed specs. The user approves this plan once and pastes a single /goal command. From this point, the system autonomously executes each phase sequentially, incorporating a robust 3-strike self-healing recovery mechanism for failures (retry, fix-spec, handoff). Each phase includes verification checks, and learnings are written back to memory, enabling smarter future runs. A final audit re-verifies against the original plan before completion.

Quick Start & Requirements

• Claude Code: Install via /plugin marketplace add https://github.com/robzilla1738/supergoal.git, followed by /plugin install supergoal@supergoal and /reload-plugins. Manual installation involves cloning to ~/.claude/skills and running /reload-plugins.
• Codex CLI: Manual installation requires cloning to ~/.codex/skills and restarting Codex.
• Prerequisites: Claude Code or Codex CLI with the built-in /goal command. GitHub SSH keys may be required for shorthand cloning; otherwise, use the HTTPS URL.

Highlighted Details

• Adaptive Phase Count: The number of execution phases is dynamically determined by task complexity, not a fixed limit.
• Memory Integration: Leverages a memory system for preloading existing knowledge and writing back new learnings to improve subsequent task executions.
• 3-Strike Self-Healing: Built-in failure recovery includes auto-retry, execution of a generated fix-spec, and eventual handoff to the user if issues persist.
• Autonomous Execution Loop: Once initiated, the system manages phase execution, verification, and memory updates without further user intervention.
• Deep Planning & Self-Critique: The planning stage includes recon, risk analysis, and a self-critique pass to refine the plan before user approval.
• Comprehensive Audit: A final audit verifies deliverables against the baseline working tree and re-runs mandatory commands to catch cross-phase regressions.
• Artifact Isolation: Each run creates an isolated directory under .supergoal/ to prevent conflicts.

Maintenance & Community

The project is at version v0.7.0, with a CHANGELOG.md available for release notes. Specific community links or contributor details are not detailed in the provided README.

Licensing & Compatibility

• License: MIT.
• Compatibility: Designed for Claude Code and Codex CLI. The MIT license permits broad commercial use and integration.

Limitations & Caveats

Running multiple /goal executions concurrently within the same working tree is unsafe; users should employ separate git worktrees for true parallel development. If subjective UI/UX criteria constitute over 30% of verification checks, the final completion message may include a warning banner. Persistent gaps or repeated failures will result in a process handoff to the user.