<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> ADHD addresses premature convergence in autoregressive reasoning for coding agents by enabling parallel divergent ideation. It targets engineers and researchers working on creative, interdisciplinary, or complex design problems, offering a method to escape local minima and surface novel solutions.

How It Works

ADHD employs a two-phase architecture: divergence and focus. During divergence, it spawns multiple isolated reasoning processes, each exploring the problem from a unique "cognitive frame" (vantage point) without shared context. This prevents anchoring. The subsequent focus phase uses a separate critic pass to score, cluster, prune invalid ideas ("traps"), and deepen promising survivors. This separation of generation and evaluation, coupled with diverse framing, aims to uncover non-obvious, viable solutions.

Quick Start & Requirements

• Installation: Available as an agent skill (npx skills add UditAkhourii/adhd), a Node/TS library (npm install adhd-agent), or a CLI (npm install -g adhd-agent).
• Prerequisites: Requires npx or npm. Authentication typically relies on ANTHROPIC_API_KEY environment variable or inherited agent auth.
• Setup: Installation is a single command. Default runs involve ~10 LLM calls, yielding 30-90s latency.
• Links: Preprint: adhdstack.github.io.

Highlighted Details

• Multi-Modal: Integrates as a skill for ~50 agents (Claude Code, Cursor, etc.), a standalone CLI, or a programmatic library.
• Cognitive Frames: Utilizes 15 built-in frames (e.g., Hardware engineer, Regulator, Biology, Game design) to re-contextualize problems and drive divergent thinking.
• Generator-Critic Split: Enforces a strict separation between idea generation (isolated, no evaluation) and idea convergence (scoring, pruning, deepening) via distinct LLM calls and system prompts.
• Evaluation Suite: Includes a reproducible benchmark (npm run evals) demonstrating significant improvements over single-shot baselines in breadth, novelty, and trap detection.

Maintenance & Community

• Maintainer: Udit Akhouri (@akhouriudit).
• Contact: researchudit@gmail.com, LinkedIn, X/Twitter.
• Collaboration: Open to research and applied-AI collaborations. No community chat links provided.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

• Use Cases: Not suitable for simple lookups, known bug fixes, or tasks with single, easily discoverable answers. Avoid for high-frequency, low-latency operations.
• Performance: Default runs are 5-10x slower and costlier than single-shot LLM calls.
• Development: Future work includes recursive deepening and cross-LLM support. The current eval suite is local-only.