Codex Autoresearch
A Codex plugin for measured experiments in a repo: a baseline metric, tested small changes, and a patch with evidence
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Runs the configured benchmark and check commands
- Can create commits and clean up experiment paths within scope
Install
Manual install
codex plugin marketplace add TheGreenCedar/AgentPluginMarketplace --ref mainRegisters the marketplace if your Codex build supports it from the terminal. Then install the plugin via /plugins.
This is third-party code. Review the repository files before installing.
What it does
The plugin runs short code-improvement loops inside Codex against a measurable goal. You set the workload, correctness checks, an edit boundary and a time budget; it records a baseline metric and tries small changes. Each step is compared to the baseline and logged as a baseline, a keep, a discard or a failure. The run leaves a reviewable patch on a branch with its evidence, and every step shows up on a local read-only dashboard. It fits work like faster tests, lower build time, less memory or lower latency.
Who it is for. For developers and engineers who improve code against a numeric metric and want reproducibility.
Good fit when
- You have a measurable target and many candidate changes to try
- You want faster tests or builds or less memory against a stable benchmark
- You need a patch with a log of attempts and evidence
Not a fit when
- The task is one-off or a matter of taste rather than a number
- The benchmark is too slow and noisy to gain from repeated measurement
Example request
/goal @Codex Autoresearch run a measured loop to speed up the parser. Benchmark: node bench/parser.mjs. Metric: seconds, lower is betterLimitations
It runs inside Codex, so it needs Codex access. The benchmark must print a metric line like METRIC seconds=12.34. The plugin can create a commit within the configured scope and clean up experiment paths, and finalization starts with a read-only preview. Session state lives in the project, under .git/autoresearch in a git repo.
How to disable. Uninstall the codex-autoresearch plugin through Codex plugin management.
Security check
- Runs the configured benchmark and check commands
- Can create commits and clean up experiment paths within scope
README in short
The README describes a Codex plugin for reproducible experiments on local code. You accept a contract of goal, checks, boundary and budget; the plugin then measures a baseline, tries changes, keeps a log and leaves a reviewable patch. Version 3.0 also supports bounded investigations where you know the outcome but not the method. It installs via the /plugins picker or a marketplace registration and ships an optional read-only dashboard. Apache-2.0 licensed.
FAQ
What must the benchmark print?
At least one metric line, for example METRIC seconds=12.34. It decides whether the result moved the right way.
Can the plugin change git?
Yes, within the configured scope: a keep can create a commit, and finalization starts with a read-only preview.
Related
A skills library that gives coding agents a development process: brainstorming, planning, TDD, subagents and code review
Skills for real engineers by Matt Pocock
Skills For Real Engineers
Small composable skills for engineering with agents: plan grilling, TDD, bug diagnosis, code review and architecture
GitHub toolkit for spec-driven development: the specify CLI adds agent commands and skills to a project, from principles to implementation
Reference MCP servers
Model Context Protocol servers
Official reference MCP servers: Filesystem, Fetch, Git, Memory, Sequential Thinking, Time and Everything