Create Repo Agent

Purpose

Build repo agents that can run unattended without granting the model broad write credentials, arbitrary shell, or uncontrolled network access. The default architecture is a read-only audit job that produces a validated bundle plus a separate publisher job that owns GitHub writes.

Use this skill together with the domain skill for the files the agent will maintain. For example, a pricing agent must also use add-model-price.

Required Reading

For every repo agent task, read these references before designing or editing:

1. references/security-standards.md
2. references/workflow-blueprint.md when implementing or changing a GitHub Actions workflow
3. references/review-checklist.md before final review or PR publication

Workflow

1. Define the exact maintenance objective, allowed files, external sources, expected no-change behavior, and PR ownership.
2. Choose the least-capable runtime: prefer a scheduled/manual GitHub Action with read-only repository checkout and no write credentials in the LLM step.
3. Encode the prompt with explicit allowed edit surfaces, hard constraints, source-evidence requirements, and structured output.
4. Give the agent only scoped file tools, domain-scoped fetch tools, and exact deterministic validator commands.
5. Validate the diff independently of the agent, including untracked files, path allowlists, git diff --check, line-count limits, and domain-specific validators.
6. Publish from a separate job or step after validation, using a bot credential only for branch push and PR create/update.
7. If self-improvement is allowed, constrain it to named workflow or skill-reference files and require security invariants to remain unchanged.
8. Run agent setup checks when .agents/** changes, then publish a normal human-reviewable PR.

Non-Negotiables

• Never expose a write-capable GitHub token, PAT, GitHub App token, OIDC token, SSH key, cloud credential, or package-publishing token to the LLM agent step.
• Never rely on prompt instructions as the only security boundary. Enforce file and command limits outside the agent.
• Never stage a directory wholesale. Stage only the validated file list.
• Never ignore untracked files in diff validation.
• Never let self-improvement bypass the same diff allowlist and human PR review as normal edits.
• Never grant arbitrary Bash, curl, wget, gh, git push, package-manager, interpreter, environment-dump, or process-inspection tools to the LLM agent.
• Never add id-token: write unless the agent truly needs OIDC and the trust relationship is reviewed explicitly.