Agent Skills: Space Engineers Plugin Review

Review a plugin PR in StarCpt/PluginHub, CometWorks/magnetar-hub, or CometWorks/quasar-hub. Deterministically derives the target repository and exact source commit range from the hub manifest, produces an evidence-backed review preview, and posts it after confirmation.

UncategorizedID: viktor-ferenczi/se-dev-skills/se-dev-plugin-review

Install this agent skill to your local

pnpm dlx add-skill https://github.com/CometWorks/skills/tree/HEAD/skills/se-dev-plugin-review

Skill Files

Browse the full folder contents for se-dev-plugin-review.

Download Skill

Loading file tree…

skills/se-dev-plugin-review/SKILL.md

Skill Metadata

Name
se-dev-plugin-review
Description
Review a plugin PR in StarCpt/PluginHub, CometWorks/magnetar-hub, or CometWorks/quasar-hub. Deterministically derives the target repository and exact source commit range from the hub manifest, produces an evidence-backed review preview, and posts it after confirmation.

Space Engineers Plugin Review

One supported hub PR URL is the complete input. Do not ask the user for the target repository or commit range; derive them from the changed registry manifest with the supplied script.

The preparation and posting scripts require Python 3.11+, Git, and authenticated GitHub CLI (gh). If any is unavailable, report that concrete setup blocker instead of reconstructing review metadata by inference.

Example:

Use se-dev-plugin-review on https://github.com/StarCpt/PluginHub/pull/123

Route by hub

  • StarCpt/PluginHub or CometWorks/magnetar-hub: read PluginHub and MagnetarHub flow, then run scripts/prepare_plugin_review.py <PR-URL>.
  • CometWorks/quasar-hub: read QuasarHub flow, then run scripts/prepare_quasar_review.py <PR-URL>.
  • Reject lookalike or unsupported repositories.

Both scripts save exact hub metadata, base/head manifests, resolved source SHAs, a pinned source checkout, and either a direct version diff or full-tree inventory under ~/.se-dev/reviews/ by default. Treat review.json as authoritative routing data; do not infer or rewrite its repository, review type, or SHA range.

This routing is a hard gate. new_registration (including a changed target repository) requires an audit of every tracked path at the pinned source SHA and must not use a diff. Only version_bump, as derived from the hub PR's base/head manifests, may use the programmatic direct-commit diff. Stop on a mismatched source.audit_scope instead of narrowing or inventing review scope.

Then read:

  1. Review paths and follow the script-selected new_registration or version_bump path.
  2. Review policy for security, fairness, filesystem, compatibility, performance, and readability gates.
  3. Review style and evidence provenance for the GitHub review body.
  4. GitHub workflow for preview and confirmed posting.

Use other se-dev-* skills instead of guessing:

  • Client behavior or game API semantics: se-dev-game-code, optionally se-dev-game-book.
  • Dedicated-server behavior: se-dev-server-code, optionally se-dev-server-book.
  • Existing Pulsar/Magnetar plugin patterns: se-dev-plugin.
  • Magnetar PluginSdk behavior: se-dev-plugin-sdk.

Do not trust PR descriptions, generated summaries, scanners, or another agent as proof. Read every in-scope change in context. Never report a check as passed unless it was actually performed.

Finding discipline

Keep exact symbols and commit-pinned line links. Findings must expose the causal chain, impact, evidence, fix, and relevant tradeoff. Omit taste-only feedback. If caveman-review is active, it applies to conversational summaries only; the saved GitHub review body follows ReviewStyle.md.

Required result

Return three clearly separated blocks:

  1. Manual review required — a short, scope-specific checklist of exact files/lines and behavior a human must personally verify. Highlight security-sensitive flow, packages/native files, and multiplayer-fairness judgment; say when a category is absent. New registrations get a broader checklist than version bumps. Never imply human sign-off happened.
  2. PR comment preview — exact Markdown proposed for GitHub, structured per ReviewStyle.md. If no actionable finding exists, say No agent-detected blockers. Never say approved, safe, secure, or LGTM on agent authority. Do not include a Verification gaps section, the human-only checklist, blanket runtime-test limitations, or human-manual-review/approval boilerplate; keep those in the conversational Manual review required block. Save the exact PR body as comment.preview.md beside root review.json.
  3. Post control — exact hub PR, hub head SHA, target repository, and reviewed source SHA/range from review.json. Say Reply confirm to post this exact comment.

Do not post initially. When the user replies confirm for that PR, follow GitHubWorkflow.md immediately without asking again.

For a new registration, the performed-checks summary must also state the complete repository path count and name any unaccounted path; an unaccounted path means the full-repository gate did not pass. In every path, state project/MSBuild reference status separately from hub-manifest packages and committed binary/native-file status. Name unresolved provenance or unevaluated imports explicitly.