Guide · Agent skill
How to use the Meta-Apply Patch Lander agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Meta-Apply is an agent skill for the ARIS research-automation skill corpus. It is the only skill that can turn a staged self-modification patch, proposed by meta-optimize, into a real change to the skill files, and it does so only after a fresh cross-model review and your explicit approval. This guide explains what the skill does, how to install it, how to trigger it, and what it needs before you run it.
Meta-Apply Patch Landerwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download and unzip the skill. Download the meta-apply ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder. Alternatively, upload it under Settings → Capabilities → Skills in the Claude apps.
- 2Check the requirements. Confirm the Codex MCP server, the provenance.py helper, and Python 3 are available. Make sure staged patches exist under .aris/meta/pending/ from a /meta-optimize run.
- 3Back up the repository. Make sure the repository is backed up or under version control, because the skill edits skill files directly.
- 4Read the meta-optimize report. Review each staged patch and its rationale in the REPORT before deciding which ones to land.
- 5Run meta-apply with your choices. Type /meta-apply followed by patch numbers or all. The skill checks the patches you named and no others.
- 6Review the landing report. Read which patches were LANDED or REFUSED and why. Keep the backup paths in case you need to revert.
What an agent skill is and what meta-apply makes the agent do
An agent skill is a folder containing a SKILL.md file that tells an AI coding agent how to handle a particular kind of task when it is triggered. Meta-apply is one such skill. Its job is to land patches: take candidate changes that other ARIS skills have staged under .aris/meta/pending/ and apply them to the skill corpus.
The skill is built around a privilege boundary. Producer skills such as meta-optimize are read-only and can only stage candidate patches. Meta-apply is the only place a staged patch becomes a real change. Splitting propose from land across two skills with different tool permissions means a loop cannot apply its own patch.
Install meta-apply with a ZIP download or the Claude apps
Download the ZIP for skills/meta-apply from this page. The skill ships as a single file, SKILL.md.
For Claude Code, unzip the folder into ~/.claude/skills/ to make it available for all your projects, or into a project's .claude/skills/ folder to limit it to that project.
In the Claude apps, you can upload the skill instead under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project
Meet the requirements before you run it
Meta-apply depends on several pieces that must already be in place. Without them, it will stop or refuse to land anything.
- Codex MCP server (mcp__codex__codex) for the cross-model jury
- The provenance.py helper from the ARIS tools folder
- Python 3, used to run provenance stamping
- Staged patches under .aris/meta/pending/, produced by /meta-optimize
Trigger the skill with natural requests
The skill is human-invoked only. It runs when you explicitly type /meta-apply after reading the report from a /meta-optimize run. The source also lists these phrases as triggers: "meta apply", "land the staged patches", and "应用优化".
You choose which patches to land. The argument is a patch number or all, so /meta-apply 1,3 lands patches 1 and 3, and /meta-apply all lands every staged patch. If you name nothing, the skill applies nothing.
- /meta-apply 1,3
- /meta-apply all
- land the staged patches
/meta-apply 1,3Understand what happens to each patch
For every patch you name, the skill runs its own review at landing time. It does not trust a verdict written by the producer. A fresh Codex review judges the staged diff and its target file, and any KILL verdict refuses the patch. You cannot override a KILL.
Patches that pass are backed up, applied with Edit or Write, stamped with a provenance record, and logged to .aris/meta/optimizations.jsonl. The report then lists each patch as LANDED or REFUSED with a reason.
- Reviewer and author must come from different model families; otherwise the patch is refused.
- Backups are written to .aris/meta/backups/<date>/<target>.
- Landed patches are removed from .aris/meta/pending/.
Safety notes: scripts, network and edits to your files
Meta-apply runs scripts and uses the network. It calls the Codex MCP server for review and runs python3 for provenance stamping. It edits skill files directly, so the skill can change the files it is meant to change.
Only use it on a repository you have backed up or keep under version control. A landed patch can be reverted from its backup, and you should test the changed skill on your next run.
- Back up or version-control the repository before landing anything.
- Review the REPORT from meta-optimize before naming patches.
- Expect network calls to the reviewer during landing.
Tips for getting reliable results
Read the producer's REPORT first and land only the patches you understand. The producer's pre-screen is advisory and does not decide landing. Only the fresh review verdict counts.
Keep the stamp in perspective. It records that a change passed a process, a cross-model review plus your landing approval. It does not prove the change is correct, and reviewer judgment can be wrong.
- Land a few patches at a time rather than all at once.
- Check the backup path in the report if you need to roll back.
- Treat the provenance stamp as a receipt, not a guarantee of correctness.
What people use it for
- A maintainer of the ARIS skill corpus reviews a meta-optimize proposal and lands only the patches they agree with.
- A maintainer who needs an audit trail checks the optimizations log to see which model authored a change and which model reviewed it.
- A developer rolls back a landed change by restoring the file from its backup under .aris/meta/backups/.
Frequently asked questions
How do I install the meta-apply skill?
Download the ZIP for skills/meta-apply from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
What do I need to run meta-apply?
You need the Codex MCP server (mcp__codex__codex) for the cross-model jury, the provenance.py helper from the ARIS tools folder, Python 3, and staged patches under .aris/meta/pending/ produced by /meta-optimize. Without staged patches there is nothing to land.
Is meta-apply free, and what is its licence?
The skill is listed under the MIT licence. This page does not state any price, so check the directory listing for cost details.
What are the alternatives to meta-apply?
This page does not list alternatives. Meta-apply is specific to landing staged patches from meta-optimize in the ARIS skill corpus, so it only applies to that workflow.
Why did my patch get refused?
A patch is refused if the fresh cross-model review returns KILL, if the author and reviewer come from the same model family, or if the provenance check fails. The report gives a one-line reason for each refusal.
Does the provenance stamp prove a change is correct?
No. The stamp is a process receipt that records who authored the change, who reviewed it at landing, and a content hash. It does not claim the change is correct, and reviewer judgment can be wrong.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Meta-apply suits developers who maintain the ARIS skill corpus, run meta-optimize, and want to review each self-modification before it lands with an audit trail. It is not useful for ordinary research or writing work, and it is risky on any repository you have not backed up.
Download the skill