Claude-skill-registry changeset-validation

Validate changesets in openai-agents-js using LLM judgment against git diffs (including uncommitted local changes). Use when packages/ or .changeset/ are modified, or when verifying PR changeset compliance and bump level.

install

source · Clone the upstream repo

git clone https://github.com/majiayu000/claude-skill-registry

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/majiayu000/claude-skill-registry "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/data/changeset-validation" ~/.claude/skills/majiayu000-claude-skill-registry-changeset-validation && rm -rf "$T"

manifest: skills/data/changeset-validation/SKILL.md

Changeset Validation

Overview

This skill validates whether changesets correctly reflect package changes and follow the repository rules. It relies on the shared prompt in

references/validation-prompt.md

so local Codex reviews and GitHub Actions share the same logic. Experimental or preview-only feature additions that are explicitly labeled as such in the diff may remain a patch bump when they do not change existing behavior. Major bumps are only allowed after the first major release; before that, do not use major bumps for feature-level changes.

Quick start

Local (Codex-driven):

Run:

pnpm changeset:validate-prompt

Apply the rules from
```
references/validation-prompt.md
```
to the generated prompt.
Respond with a JSON verdict containing ok/errors/warnings/required_bump (English-only strings).

CI (Codex Action):

Run:

pnpm changeset:validate-prompt -- --ci --output .github/codex/prompts/changeset-validation.generated.md

Use
```
openai/codex-action
```
with the generated prompt and JSON schema to get a structured verdict.

Workflow

Generate the prompt context via
```
pnpm changeset:validate-prompt
```
.
Apply the rules in
```
references/validation-prompt.md
```
to judge correctness.
Provide a clear verdict and required bump (patch/minor/major/none).
If the changeset needs edits, update it and re-run the validation.

Shared source of truth

Keep the prompt file as the single source of validation rules.
Keep the script lightweight: it should only gather context and emit the prompt.

Resources

```
references/validation-prompt.md
```