Propel Documentation
A structured constraint framework that turns unconstrained LLM output into precise, paper-aligned research code.
Getting Started
New to Propel? Start here to understand what it does and how to set it up.
Installation & Setup
Install Propel, initialize it in your project, and run your first workflow in under 5 minutes.
Core Principles
The five non-negotiable principles injected into every session — the mindset beneath all mechanisms.
The Pipeline
Understand the end-to-end workflow: gates, questioners, and phase transitions.
Pipeline Overview
The full seven-stage workflow from intake to retrospective — how each phase connects and why ordering matters.
Gates & Questioners
Deep dive into Gate 0–4 and Questioners Q0–Q1. What each checkpoint asks, why, and what format.
Modes
Not every session needs the full pipeline. Each mode filters which skills and gates are active.
Researcher Mode
Literature review, investigation, deep research. Understanding the problem space before building anything.
Engineer Mode
Full pipeline with all gates and auditors. The default mode for implementing paper-derived code.
Debugger Mode
Root-cause analysis with evidence-backed diagnosis. Classifies code bugs vs. design issues vs. config problems.
Trainer Mode
Launch training runs, monitor progress, fix CUDA/OOM/path errors. Code is ready — time to train.
Skills
17 specialized skills organized by workflow phase. Each activates on specific triggers and guides the agent through structured processes.
Investigation
Scaffolds scratch/ directories with living READMEs that persist across sessions.
Research Design
Paper-to-code mapping, regression risk assessment, and explicit design decisions before implementation.
Systematic Debugging
Diagnosis-first methodology with 3-strike limit. Never apply a fix without presenting the root cause.
All Skills →
Complete reference of all 17 skills: triggers, active modes, and detailed workflow descriptions.
Agents (Auditors)
Domain-specific auditors that auto-dispatch after code changes. They run as subagents with separate context windows and read-only access.
Silent Bug Detector
Scans for 11 categories of silent failures — broadcasting bugs, wrong reductions, detached gradients, and more.
Paper Alignment Auditor
Cross-references code against paper equations and algorithms to catch implementation mismatches.
JAX Logic Auditor
Traces tensor shapes through JAX transforms (vmap, jit, scan) to catch shape mismatches.
All Agents →
Complete reference of all 8 auditors: dispatch rules, auto-trigger conditions, and output formats.