প্রকল্প সম্পর্কে

wicked-garden is a specialized toolkit for coding agents like Claude Code, Codex, and Cursor, augmenting planning and swarm capabilities. Its core philosophy, 'done is re-derived, not asserted,' ensures task completion is verified through independent evidence rather than self-reporting. Key capabilities include: 1. **Evidence-Gated Verification**: Re-runs proofs to reject false 'tests pass' claims, ensuring no vacuous green checkmarks. 2. **Code Graph Analysis**: Identifies 'injected edges' (e.g., event-to-consumer or command-to-agent links) missed by grep, enabling lineage tracking. 3. **Cross-Session Memory**: Maintains context across long-term projects via `wicked-estate`, helping agents maintain context. 4. **Multi-Model Review Panels**: Convenes 'councils' of AI models (e.g., Antigravity, Codex) for second opinions on architectural decisions. 5. **Repo-Specific Playbooks**: Loads generated playbooks based on the current HEAD, reducing errors in complex codebases. 6. **Deterministic Refactoring**: Renames as graph operations to minimize errors in complex codebases. wicked-garden operates on a 'work-shape' model, applying rigor based on task type (e.g., typo fix vs. migration cutover). Extensible with domain packs, requiring Python 3.9+ and Node.js with `wicked-vault` for evidence gates. Integrates with CLIs and compiles verification gates into repositories, ensuring portable enforcement independent of installation.