monarchjuno/tradingcodex

TradingCodex is a framework for doing investment research and portfolio work through OpenAI's Codex coding agent. The core problem it addresses is that complex investment workflows done through chat tend to disappear when the conversation ends. This project gives Codex a durable backing system: research notes saved as files, a local database that tracks orders and approvals, a web dashboard for reviewing state, and a controlled boundary layer that gates any action that could actually move money. The system is organized around roles. A head manager agent receives a request like "analyze Apple with public equity research and portfolio review" and then divides that work among specialist agents covering research, valuation, risk, and portfolio fit. Each specialist produces a bounded artifact, a markdown document with sources and timestamps, which the next role reads rather than redoing the same work. This handoff model is designed to avoid the situation where every agent independently re-derives the same context. All the parts that touch state, such as tracking whether an order is pending approval or what the current portfolio looks like, run through a local Django web application. That application also exposes a browser dashboard at a local port where you can browse agents, research documents, order history, and policy settings without digging through chat transcripts. The same logic that the dashboard calls is what the command-line tools and the AI agent layer call, so there are no separate execution paths that might diverge. The project is explicit that it is not an autonomous trading bot. There are no live broker connections in the current release. Order submission goes through a paper or stub path, meaning the system records what an order would look like and requires a human approval step before anything real could happen. The execution boundary, implemented as a set of typed tools that agents can call, enforces role permissions and policy checks on every action. Installing it generates a new Codex workspace folder on your machine with all the config, role prompts, and wrapper scripts pre-populated. A doctor command checks that everything is wired correctly after setup. The README is detailed and aimed at developers comfortable with Python tools and the Codex agent environment.