TheGreenCedar/codex-autoresearch

A codex plugin for running optimization loops inside a codebase. It is useful when you have a measurable target and many possible changes to try: test runtime, build speed, bundle size, model loss, Lighthouse scores, memory use, query latency, or any other metric you can print from a script.

What it solves

Codex Autoresearch automates the process of improving local code through repeatable experiments. It eliminates the manual trial-and-error cycle of making small code changes and measuring their impact on specific performance or resource metrics, such as runtime, memory usage, or build speed.

How it works

Users provide a "contract" consisting of a workload (benchmark), correctness checks (tests), an edit boundary (scope), and a time/attempt budget. The tool then executes a loop: it establishes a baseline measurement, proposes and evaluates small changes, and logs the results. If a change improves the primary metric without breaking correctness checks, it is kept; otherwise, it is discarded. The process concludes with a reviewable patch and an evidence receipt showing the results of the experiments.

Who it’s for

Developers who need to optimize specific parts of their codebase for performance or resource efficiency and have a stable, measurable benchmark to test against.

Highlights

  • Contract-based execution: Requires explicit approval of goals, metrics, and boundaries before running.
  • Automated iteration: Handles the setup, measurement, and logging of multiple candidates automatically.
  • Evidence-based results: Provides a final patch backed by a detailed record of measurements and outcomes.
  • Integrated Dashboard: Offers a read-only visual readout of improvements, passes, and blockers.
  • Git integration: Manages state within the repository and can create commits for successful changes.

Related

  • Project
    leo-lilinxiao/codex-autoresearch

    Codex‑autoresearch is a Git‑backed Codex skill that runs an autonomous loop: propose a single code change, commit it, measure a numeric metric (e.g., test failures), keep the change if it improves the metric and passes optional guards, otherwise revert. It repeats until a user‑specified target is reached, storing an immutable event log and HTML report. Supports foreground or background execution, works with any measurable outcome, and enforces strict safety (all trials are real commits, failures abort with clear errors).

  • Project
    uditgoenka/autoresearch

    An autonomous improvement engine for AI coding agents that uses a loop of modification, mechanical verification, and git-based rollbacks to optimize code and systems.

  • Project
    drivelineresearch/autoresearch-claude-code

    autoresearch‑claude‑code is a Python framework that runs budget‑constrained auto‑research loops with Claude Code or Codex agents. You define a goal, a deterministic scorer, and files the agent may edit; the system creates a Git branch, measures each change, and automatically commits only statistically‑significant improvements while respecting run‑time and turn limits.

  • Project
    davebcn87/pi-autoresearch

    pi‑autoresearch is a pi‑coding‑agent extension that lets the AI run autonomous optimisation loops: it creates a session, repeatedly executes a user‑defined benchmark command, logs the metric, auto‑commits good changes, and provides a live dashboard. After the loop, it can split successful runs into reviewable git branches. The tool is domain‑agnostic, configurable, and supports optional hooks and confidence scoring.

  • Project
    chAng-L19/codex-redteam-mode

    A durable red-team runtime for Codex App and CLI that automates security workflows using goal contracts and evidence-based verification to prevent session interruptions and false completions.