Autoresearch Wrapper — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Autoresearch Wrapper (Plugin) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
Aggregate score unchanged between these scans.
The primary manifest — the file an agent reads to learn what this artifact does.
<div align="center">
Autoresearch Wrapper Banner
Dependency-aware optimization engine for any codebase.
Scan, plan, isolate, measure, repeat — all in Git worktrees, hands-free.
Inspired by Karpathy's autoresearch — constraint + mechanical metric + autonomous iteration = compounding gains.
<br>
"Just type /autoresearch-wrapper → everything's managed"
<br>
Install · Usage · Commands · Docs · 简体中文
</div>
Recently, I've been asking Codex to write autoresearch wrappers for the modules that I want to optimize. There's a concrete use case where I want to add a feature or module, and I have several candidates. But each time you add one, you need to optimize all the other related parts to get a sense of how helpful the module is (autoresearch-wrapper-create). That's when I wanted to implement a SKILL for this, and I thought that at the very beginning, before optimizing a system, we should understand the dependencies. So here it comes: a Dependency-aware Autoresearch Engine that wraps your modules, scripts, and more.
SKILL.md documents every command, so the repo root can be installed as one skill.resources and reference commands and the .autoresearch-wrapper/plans/ files are gone (the CLI tells you what to do instead).create / delete are now fair comparisons: every approach (or the deletion) and a no-change control arm are tuned with the same budget, then ranked by their tuned ceiling with compare.finalize merges the winner back and cleans up worktrees.ln -s /path/to/autoresearch-wrapper/skills/* ~/.agents/skills/ gives you every command as its own skill; the old repo-root install still works as one comprehensive skill.<details open> <summary><b>Claude Code (recommended)</b></summary>
/plugin marketplace add HenryCai11/autoresearch-wrapper
/plugin install autoresearch-wrapperTo update to the latest version:
/plugin marketplace update autoresearch-wrapper
/plugin update autoresearch-wrapperIf you installed an older 0.0.2 marketplace build before the plugin packaging changes, run:
/plugin marketplace update autoresearch-wrapper
/plugin remove autoresearch-wrapper
/plugin install autoresearch-wrapperTo uninstall:
/plugin remove autoresearch-wrapperOr install manually:
git clone https://github.com/HenryCai11/autoresearch-wrapper.gitClaude Code discovers skills from .claude/skills/ automatically. To use them in all your projects, symlink the skill folders into your user skills directory:
mkdir -p ~/.claude/skills
ln -s /path/to/autoresearch-wrapper/skills/* ~/.claude/skills/The marketplace package lives under plugins/autoresearch-wrapper/ and is generated from skills/.
</details>
<details> <summary><b>Codex</b></summary>
Codex reads the same skill folders. Inside this repo they are picked up from .agents/skills/. To use them everywhere:
mkdir -p ~/.agents/skills
ln -s /path/to/autoresearch-wrapper/skills/* ~/.agents/skills/Invoke skills with $autoresearch-wrapper, $autoresearch-wrapper-run, and so on (or pick from /skills). Restart Codex after installing.
Alternatively, install the repo root as one comprehensive skill (ln -s /path/to/autoresearch-wrapper ~/.agents/skills/autoresearch-wrapper); the root SKILL.md covers every command. The older ~/.codex/skills/autoresearch-wrapper install keeps working the same way. Pick one style: both provide a skill named autoresearch-wrapper.
</details>
<details> <summary><b>CLI only</b></summary>
No installation needed — run directly:
python3 scripts/autoresearch_wrapper.py scan</details>
Just run
/autoresearch-wrapperThis starts a wizard that begins with a compact core-functionality scan, then guides you step-by-step through target selection and run setup. If you want the full repo-wide listing, use python3 scripts/autoresearch_wrapper.py wizard --full-summary.
An example of the wizard for optimizing Recursive Language Models
example_rlm
One skill set serves both tools: invoke a skill as /<skill> in Claude Code or $<skill> in Codex.
| Skill | CLI | What it does |
|---|---|---|
autoresearch-wrapper | wizard / wrap | Run the end-to-end wizard or wrap a script |
autoresearch-wrapper-status | status | Show state and readiness |
autoresearch-wrapper-run | run | Start or resume a run |
autoresearch-wrapper-flow | flow | Metric history and plot |
autoresearch-wrapper-create | create | Compare feature approaches, each tuned vs. a tuned control |
autoresearch-wrapper-delete | delete | Compare deleting a part vs. keeping it, both tuned |
autoresearch-wrapper-monitor | monitor | Progress snapshots (the CLI poller is for terminals) |
Additional CLI subcommands: scan, configure, allocate, evaluate, record, arm-ready, compare, finalize, preset-metric
When a run ends, finalize --run-id <id> --merge squash-merges the best candidate into your checkout (staged for review, not committed) and removes the run's worktrees.
python3 -m unittest -q| Doc | Contents |
|---|---|
| DETAILS.md | Feature details, dependency graph notes, generated layout, typical workflow |
| CORE_FEATURES.md | Feature checklist (Chinese) |
| CONTRIBUTING.md | Development workflow, branch layout, verification |
| THIRD_PARTY_NOTICES.md | Upstream attribution |
I'm inspired by several autoresearch projects. They are Karpathy's autoresearch, and uditgoenka's autoresearch. Thanks!
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.