ADHD is a skill for coding agents that implements tree-of-thought reasoning with pruning, designed to combat premature convergence in autoregressive models by spawning parallel divergent thought processes under different cognitive frames.
The tool addresses a fundamental architectural problem in chain-of-thought reasoning: linear approaches anchor on initial outputs, and even tree-of-thought methods maintain shared context across branches, perpetuating early biases. ADHD solves this by spawning N isolated reasoning processes under deliberately distorted cognitive frames with zero shared context during divergence, then running a separate critic pass to score, cluster, prune traps, and deepen the survivors. This approach treats convergence bias as an architectural issue rather than a prompting problem.
The tool is built for design decisions, fuzzy debugging, naming, API surface design, strategy work, and any prompt asking for multiple approaches or solutions. It suits creative and interdisciplinary work where exploring a wide solution space matters more than fast convergence. The README demonstrates this through a concrete eval problem on CLI timeout strategy, where a baseline single-shot approach produces textbook patterns and sensible but conventional recommendations, while ADHD surfaces 30+ ideas across multiple cognitive frames including economic-incentive, async-control-surface, gamification, and redundancy-race clusters, identifying traps and user agency considerations the baseline misses.
The project maintains an active Discord community focused on frame design, evaluation problems, and trap-hunting research. Development is coordinated through community contribution channels for frame contributions, integrations, and adopter onboarding. The tool is built on the Claude and Codex Agent SDKs in TypeScript and includes benchmark results with full transcripts documenting comparative performance on reasoning tasks.
