skills.rest
langfuse
langfuseOFFICIAL
Explore all skills by this author →
Website𝕏@langfuseGitHub

create-repo-agent

Design and harden LLM-powered GitHub Actions agents with read-only audit and separate publish jobs.

34.0k|3.7k|Updated May 18, 2023
Search Tags:#prompt injection#LLM agent#GitHub actions#CI security#Code review#Workflow Automation#devOps
One-click install
npx skills add https://github.com/langfuse/langfuse --skill create-repo-agent
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: create-repo-agent
Source: https://github.com/langfuse/langfuse/tree/main/.agents/skills/create-repo-agent
Command: npx skills add https://github.com/langfuse/langfuse --skill create-repo-agent

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Autonomous repo agents that run LLMs in CI can leak write credentials, execute arbitrary commands, or push unvalidated changes. This Skill provides security standards, a workflow blueprint, and a review checklist for building repo-owned agents that run unattended without granting the model broad write access.

Core Features & Use Cases

  • Two-Phase Architecture: Separates a read-only LLM audit job that produces a validated patch artifact from a publish job that owns GitHub writes and never invokes the LLM.
  • Security Standards: Enforces credential isolation, scoped tool allowlists, anchored path allowlists, staged-blob validation, and prompt-injection threat modeling.
  • Review Checklist: Provides merge-blocking criteria covering credentials, tools, diff validation, publish boundaries, and self-improvement constraints.
  • Use Case: When building a scheduled pricing-maintenance agent that fetches provider pages and opens PRs, use this Skill to design the workflow so the LLM step has no write token, no shell access, and every changed file passes independent diff validation before a bot publishes the PR.

Quick Start

Use the create-repo-agent skill to design a scheduled GitHub Actions agent that audits a config file and opens a validated PR without giving the LLM write credentials.

Frequently Asked Questions about create-repo-agent

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I build a secure LLM agent in GitHub Actions?▼

Use a two-phase architecture: a read-only audit job runs the LLM with only the model API key and produces a validated patch artifact, then a separate publish job applies the patch and pushes the branch. The LLM step never receives write tokens, shell access, or broad network tools.

How to prevent prompt injection in CI automation agents?▼

Treat all web pages, issues, and prior outputs as untrusted input, and never rely on the prompt as the only guardrail. Enforce file and command limits outside the agent with scoped tools, anchored path allowlists, and independent diff validation before publishing.

What permissions should a GitHub Actions LLM step have?▼

The LLM step should have only contents: read, checkout with persist-credentials: false, and the model API key plus a read-only github.token if required. Write credentials, PATs, OIDC tokens, and package tokens belong only in a separate publish job that does not invoke the LLM.

Can an AI agent safely update its own workflow file?▼

Self-improvement is allowed only when the workflow explicitly opts in, limited to named files like the workflow itself and skill references. Security invariants such as read-only audit jobs and path allowlists must remain unchanged, and all self-edits go through the same diff validation and human PR review.

Why does diff validation need to include untracked files?▼

An agent could create new files outside the allowlist that git diff alone would miss. Validation must combine git diff --name-only with git ls-files --others --exclude-standard, intent-to-add untracked files, and check every path against the anchored allowlist before staging.

Related Skills

Explore More Software Engineering Skills→
MING-CHUNLeeMING-CHUNLee

mvc-architecture

Implement Gateway, Data Mapper, and Domain Entity patterns in Service Oriented Architecture.

Community
Advanced
tillysosotillysoso

babysit-pr

Monitor pull request health across CI/CD platforms and triage review threads.

Community
Advanced
doctordukedoctorduke

doc-to-skill-scraper

Scrape external documentation and generate SKILL.md files with YAML frontmatter.

Official
Intermediate
whichguywhichguy

cron-llm-review-house-style

Review cron job outputs with LLM and apply house-style formatting.

Community
Advanced