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 maintainers of the ARIS auto-research harness. It takes patches that a read-only producer staged under .aris/meta/pending/ and, only when you name them, runs a fresh reviewer on each one, then applies the approved ones to the corpus. 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
- 1Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 2Meet the requirements. Make sure a reviewer backend is configured, that tools/provenance.py from the ARIS repository is available, and that Python 3 is installed.
- 3Run the producer first. Run /meta-optimize so that candidate patches are staged in .aris/meta/pending/.
- 4Read the REPORT. Review the producer's REPORT and decide which patches you accept. Do not rely on the producer's advisory pre-screen.
- 5Name the patches. Invoke the skill with the patch numbers you want, for example /meta-apply 1,3.
- 6Check the result. Read the per-patch output: LANDED with a backup path, or REFUSED with a reason. Test the changed skill on your next run.
What is an agent skill, and what does meta-apply do?
An agent skill is a folder containing a SKILL.md file that tells an AI coding agent when and how to perform a task. Meta-apply is one such skill, used in the ARIS research harness.
Its job is to be the only step that turns a proposed change to ARIS's own skills and corpus files into a real edit. Read-only skills such as /meta-optimize stage candidate patches in .aris/meta/pending/. When you run meta-apply and name patch numbers, it reviews each staged diff, refuses any patch the reviewer rejects, backs up the target file, applies the change with Edit or Write, and records a provenance stamp.
- It only handles staged ARIS self-modification patches under .aris/meta/pending/.
- It is human-invoked only. Your explicit invocation is the approval step.
- A reviewer KILL cannot be overridden by you; you can only choose among patches the reviewer passed.
Install meta-apply in Claude Code or the Claude apps
You can install the skill from the ZIP download on this page. Choose one of the methods below.
For Claude Code, unzip the download into your user skills folder ~/.claude/skills/, or into a project's .claude/skills/ folder if you want it limited to one repository.
For the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Check the requirements before you run it
Meta-apply needs a reviewer backend. The default is Codex; you can select another through .aris/reviewer.json. It also needs tools/provenance.py from the ARIS repository, which requires Python 3. If you select antigravity, grok or claude-review as the reviewer, those MCP reviewer tools must be accessible.
The skill runs scripts and uses the network, because the reviewer step calls a model.
- Codex or another configured reviewer backend
- tools/provenance.py from the ARIS repository (Python 3)
- MCP reviewer tools (antigravity, grok, claude-review) if you choose them as the backend
- Runs scripts: yes. Uses network: yes.
How to trigger the skill with natural requests
Run it after a /meta-optimize run, once you have read its REPORT. The skill recognises these phrases: "meta apply", "/meta-apply", "land the staged patches", and "应用优化".
Name the patches you want. The source supports patch numbers or 'all'. Name only the patches you have read, because 'all' applies every survivor of the review.
- /meta-apply 1,3 lands patches 1 and 3 only
- /meta-apply all lands every patch the reviewer passes (read the REPORT first)
- land the staged patches (asks for the staged set; name the numbers you want)
/meta-apply 1,3What happens during a landing
For each patch you name, the skill runs in order. Any failure means the patch is skipped and reported, never silently applied.
First it checks that you named the patch. Then it spawns a fresh reviewer on the staged diff and its target, using paths only. The reviewer answers PASS or KILL with a one-line reason. On PASS, the target is backed up to .aris/meta/backups/<date>/, the change is applied with Edit or Write, and provenance is stamped. Each landing is logged to .aris/meta/optimizations.jsonl.
The review class is recorded honestly. A same-family review uses stamp-provisional and does not authorize future auto-curation. A completed positive review with known different model families, or an applicable deterministic verifier, can use strict stamp.
- Named by you: only patches you listed
- Fresh review: PASS required; KILL means refused
- Backup: copy of the target in .aris/meta/backups/<date>/
- Apply: Edit or Write only, never Bash on the corpus
- Stamp and log: provenance receipt plus optimizations.jsonl entry
Safety notes: scripts, network and reversibility
Meta-apply can change ARIS's own skills and corpus files, so read the REPORT before landing anything. The skill is designed for a human to invoke, not for unattended or automated runs. Running 'all' without reading the REPORT can change the harness in ways you did not intend.
A landed patch can be reverted from its backup. After landing, the skill reminds you to test the changed skill on the next run.
A provenance stamp is a receipt that the change passed a review and a human approval process. It is not proof the change is correct. The stamp includes a content hash, so a later hand edit invalidates it.
- Read the REPORT before naming patch numbers
- Do not run it unattended
- Revert a landed patch using its backup in .aris/meta/backups/
- A stamp does not guarantee correctness
What people use it for
- Accepting a single reviewed fix to an ARIS skill after a /meta-optimize run
- Landing a corpus change proposed by the corpus-audit producer once you have checked its REPORT
- Keeping an audit trail that shows whether each landed change was reviewed by a same-family or a different-family model
- Reverting a landed change by restoring its file from the backup in .aris/meta/backups/
Frequently asked questions
What does the meta-apply skill do?
It lands staged patches from .aris/meta/pending/ into the ARIS corpus. It runs a fresh reviewer on each patch you name, refuses any the reviewer kills, backs up the target, applies the change with Edit or Write, and records a provenance stamp.
How do I install the meta-apply skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
Is meta-apply free, and what is its licence?
The skill is listed under the MIT licence. This page does not state prices, so check the ARIS repository for any other terms.
What do I need to run meta-apply?
You need a configured reviewer backend (Codex by default, or another set in .aris/reviewer.json), tools/provenance.py from the ARIS repository, and Python 3. MCP reviewer tools are needed only if you choose them as the backend.
Can meta-apply apply changes to my normal project code?
No. It only handles staged ARIS self-modification patches under .aris/meta/pending/. It is not meant for ordinary code changes in other projects.
Why did my patch get refused?
The fresh reviewer returned KILL with a one-line reason, and a KILL cannot be overridden. You can only land patches that the reviewer passed.
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 maintainers of the ARIS auto-research harness who have read a /meta-optimize REPORT and want to accept specific patches with an audit trail. It is a poor fit for anyone wanting to apply code changes to an ordinary project, or for unattended automation, since it is designed to be invoked by a human.
Download the skill