yolo-random

A Claude Code plugin that spends your leftover quota on real open-source.

Before your quota resets, run /yolo random. It picks one beginner-friendly issue from a curated opt-in list, reads the code in a throwaway temp clone, and submits a fix — by one of two methods you choose.

Two methods

	comment (default, non-invasive)	pr (real contribution)
What	One issue comment: root cause + ready-to-apply diff	Fork the repo + open a real pull request
Footprint	Zero — read-only temp clone, deleted after	A fork repo in your GitHub account
Pros	No fork/PR, nothing in your account	Counts as a contribution (profile credit), CI runs, mergeable
Cons	Not a PR — no profile credit, no CI, maintainer applies it	Leaves a fork in your account

Both sides have a say:

• Each repo declares which methods it accepts in CANDIDATES.md (comment, pr, or both). The skill never uses a method a repo doesn't list.
• You preset your preference: /yolo random pr or /yolo random comment. No arg → the skill shows the table above and asks once, then runs autonomously.

Principles

• One issue, one submission per run. No spam, no batch firing.
• Opt-in repos only — targets come from CANDIDATES.md.
• Non-invasive by default — comment method touches nothing but a temp clone.
• Honest — every submission discloses AI authorship and respects the repo's CONTRIBUTING.md.

Install

/plugin marketplace add spotwhale/yolo-random
/plugin install yolo-random@yolo-random

Running it passively (no per-step prompts)

The skill decides everything on its own — it never stops to ask "shall I proceed?".

Non-invasive tools are auto-approved by default. The plugin ships a PreToolUse hook that auto-allows web access and code reading (WebFetch, WebSearch, Read, Grep, Glob) with no prompts. Nothing to configure — it activates when the plugin is enabled.

Shell actions are NOT auto-approved — the temp clone, gh issue comment, rm, and any other Bash command still prompt. This is deliberate: a plugin can't safely auto-approve arbitrary shell (command chaining like git status; curl x | sh defeats any allowlist), so you opt into those once. For a fully hands-off run:

claude --dangerously-skip-permissions
> /yolo random

Zero prompts, scoped to that one session. Recommended for "burn my quota" runs.

Prefer a persistent, narrower allowlist instead of the flag? Add the skill's outward commands to your ~/.claude/settings.json:

{
  "permissions": {
    "allow": [
      "Bash(gh:*)",
      "Bash(git:*)",
      "Bash(curl:*)",
      "Bash(mktemp:*)",
      "Bash(rm -rf /tmp/*)"
    ]
  }
}

Then just /yolo random runs unattended. No plugin can grant these for you — that's a deliberate Claude Code security boundary, so you allow them once.

Submit your repo (no fork needed)

Maintainers: open an issue with the Add my repo form. Opening an issue needs no fork. Pick your preferred method (comment / pr / both) and confirm the basics; a curator reviews and adds you to CANDIDATES.md.

Curating candidates

Curators: review candidate-labeled issues and add accepted repos to CANDIDATES.md with the right Method. Only add repos that welcome outside contributions and whose CONTRIBUTING.md does not ban AI-assisted contributions. The skill re-checks each repo's rules at run time, but the list is the first gate.

Requirements

• gh CLI, authenticated (gh auth status).
• comment method needs nothing more (read-only clones of public repos).
• pr method also needs a GitHub account that can fork public repos.