Mysti – Claude, Codex, and Gemini debate your code, then synthesize

Orchestrates parallel code review sessions across Claude, Codex, and Gemini by submitting the same code snippet to each model's API simultaneously, collecting structured responses, and managing the debate flow through a coordinator pattern. Each model receives identical context and prompts designed to elicit critical analysis, then responses are aggregated for synthesis. The system handles API rate limits, timeouts, and model-specific response formatting through adapter layers.

Aggregates critique and suggestions from multiple models into a unified synthesis by parsing model-specific response formats, extracting common themes, identifying disagreements, and generating a consolidated recommendation. Uses heuristic matching or embedding-based similarity to group similar suggestions across models despite different wording, then ranks recommendations by consensus strength. The synthesis layer abstracts away model-specific quirks (Claude's verbose explanations vs Codex's concise suggestions) into a normalized output format.

Provides a command-line interface that accepts code input (via stdin, file path, or clipboard), submits it to the multi-model debate engine, and streams results back to the terminal as they arrive from each model. Uses a streaming architecture where model responses are printed incrementally rather than buffered, allowing developers to see debate progress in real-time. Handles input parsing (detecting language, extracting code blocks from markdown), output formatting (syntax highlighting, colored diff output), and result persistence (optional JSON export).

Automatically detects programming language from code snippet or file extension, extracts relevant context (function signature, class definition, imports, surrounding code), and formats code for submission to models. Uses language-specific parsers or regex patterns to identify code boundaries, strip comments/docstrings for cleaner analysis, and preserve syntax highlighting metadata. Handles polyglot inputs (mixed languages in one file) by segmenting code by language before submission.

Allows users to customize the prompts sent to each model, adjust model-specific parameters (temperature, max tokens, top-p), and define debate focus areas (security, performance, style, readability). Stores configurations in YAML or JSON files that can be version-controlled and shared across teams. Supports preset debate profiles (e.g., 'security-focused', 'performance-optimized') that adjust prompts and parameters automatically, and allows per-model customization (e.g., higher temperature for Claude to encourage creative suggestions, lower for Codex for deterministic output).

Handles API failures, rate limiting, timeouts, and model-specific response formats by implementing retry logic with exponential backoff, fallback strategies (e.g., skip a model if it times out), and response parsing that tolerates malformed output. Normalizes responses from different models into a common schema (model name, critique text, severity level, suggested fix) despite different output formats. Implements graceful degradation — if one model fails, the debate continues with the other two rather than failing entirely.