Auto Research is a cross-platform autonomous iteration engine for coding agents. It keeps the existing Claude /autoresearch command surface intact, adds a root-centered Codex skill bundle and local Codex plugin, and positions the project as a reusable workflow system instead of a single-runtime package.
Inspired by Karpathy's autoresearch. The core loop is still the same:
Modify -> Verify -> Keep or Discard -> Repeat
- Public brand renamed to Auto Research
- Repository renamed to
Maleick/AutoResearch - Root
SKILL.md,references/,scripts/, andagents/now act as the Codex-facing source of truth plugins/codex-autoresearchis generated from that root bundleplugins/autoresearchremains the Claude compatibility package with the stable/autoresearch*command family- Results logging now supports both
research-results.tsvandautoresearch-results.tsv
| Surface | Status | Entry point |
|---|---|---|
| Claude | Stable compatibility layer | /autoresearch, /autoresearch:plan, /autoresearch:debug, /autoresearch:fix, /autoresearch:learn, /autoresearch:predict, /autoresearch:scenario, /autoresearch:security, /autoresearch:ship |
| Codex | First-class local plugin | $codex-autoresearch via root skill bundle or install codex-autoresearch from the repo marketplace |
SKILL.md # Root Auto Research skill bundle for Codex
agents/openai.yaml # Codex skill metadata
references/ # Shared loop, mode, and runtime references
scripts/ # Runtime helpers, sync, and validation tooling
plugins/autoresearch/ # Claude compatibility package
plugins/codex-autoresearch/ # Codex plugin package mirrored from the root bundle
.agents/plugins/marketplace.json # Repo-scoped Codex marketplace for local install/testing
docs/ # Public site and architecture docs
wiki/ # Reference docs
claude plugin marketplace add Maleick/AutoResearch
claude plugin install autoresearch@Maleick-AutoResearchThe Claude package keeps the existing autoresearch identifiers for compatibility.
This repository includes a repo-scoped Codex marketplace at .agents/plugins/marketplace.json.
- Start Codex in this repository.
- Open
/plugins. - Choose the repo marketplace and install
codex-autoresearch.
You can also invoke the root bundle directly with $codex-autoresearch while iterating on the plugin locally.
Auto Research requires a goal, scope, and a mechanical verification command. It then:
- Baselines the current state.
- Makes one focused experiment.
- Verifies it mechanically.
- Keeps strict improvements and discards regressions.
- Records the result and continues until the stop condition is met.
Common artifacts:
| Artifact | Purpose |
|---|---|
autoresearch-state.json |
Checkpoint state for the current run |
research-results.tsv |
Primary Codex-facing iteration log |
autoresearch-results.tsv |
Claude-compatible results log alias |
autoresearch-report.md |
End-of-run report |
autoresearch-launch.json |
Background launch request |
autoresearch-memory.md |
Reusable memory for later runs |
When you change the root bundle, re-sync and validate the Codex plugin payload:
python3 scripts/sync_plugin_payload.py --repo .
python3 scripts/check_plugin_distribution.py --repo .
pytest -q tests/test_plugin_distribution.py tests/test_results_compatibility.pyThe public website URL still uses the historical /claude-autoresearch/ path during the transition, but the GitHub repository and install references now point at Maleick/AutoResearch.