Curated plugin skill descriptions should be linted to fit the skills metadata budget

Summary

Several first-party plugin skills ship long frontmatter description fields. Codex renders loaded skill descriptions into the model-visible Available skills metadata block, which is capped by the skills metadata budget. Long first-party descriptions can therefore consume a large share of that budget and contribute to truncation warnings before user/project skills have much room.

This is related to #19679 (make the budget configurable) and #19090 (truncation effects), but this issue is about catalog hygiene: curated/runtime/bundled plugin skill descriptions should be concise and CI-linted.

Environment

• Codex CLI: codex-cli 0.125.0-alpha.3
• OS: macOS
• First-party plugin cache roots inspected: openai-curated, openai-primary-runtime, openai-bundled
• Note: openai/plugins appears to be the repo referenced by several curated plugin manifests, but Issues are disabled there, so I am filing this in openai/codex.

Observed

A session with local/project skills plus first-party plugin skills produced:

Warning: Exceeded skills context budget of 2%. Loaded skill descriptions were truncated by an average of 235 characters per skill.

Local audit of only first-party plugin SKILL.md frontmatter descriptions found:

• 25 plugin skills
• 7,908 total description characters
• 19 descriptions over 220 characters

Largest examples:

775  openai-curated/hugging-face/6807e4de/skills/vision-trainer/SKILL.md
629  openai-curated/hugging-face/6807e4de/skills/llm-trainer/SKILL.md
561  openai-curated/hugging-face/6807e4de/skills/jobs/SKILL.md
429  openai-curated/gmail/6807e4de/skills/gmail/SKILL.md
393  openai-curated/hugging-face/6807e4de/skills/transformers.js/SKILL.md
387  openai-curated/github/6807e4de/skills/gh-address-comments/SKILL.md
346  openai-curated/google-calendar/6807e4de/skills/google-calendar/SKILL.md
319  openai-curated/hugging-face/6807e4de/skills/papers/SKILL.md
314  openai-curated/hugging-face/6807e4de/skills/trackio/SKILL.md
297  openai-curated/hugging-face/6807e4de/skills/community-evals/SKILL.md
293  openai-curated/google-calendar/6807e4de/skills/google-calendar-group-scheduler/SKILL.md
288  openai-curated/google-calendar/6807e4de/skills/google-calendar-meeting-prep/SKILL.md
284  openai-primary-runtime/documents/26.423.10653/skills/documents/SKILL.md
264  openai-curated/gmail/6807e4de/skills/gmail-inbox-triage/SKILL.md
260  openai-curated/google-calendar/6807e4de/skills/google-calendar-daily-brief/SKILL.md

In the same local setup, trimming only repo-owned/custom skill descriptions reduced total skill-description payload from about 15,822 chars to about 13,992 chars and stopped the warning in a fresh codex exec smoke run. That workaround is fragile, though: first-party plugin descriptions still dominate a large part of the loaded metadata block and can reintroduce the warning as plugin catalogs grow.

Expected behavior

First-party skill frontmatter descriptions should be short routing metadata, not mini documentation. Long procedure details can stay in the skill body where progressive disclosure already applies.

Suggested fix

• Add a packaging/CI lint for first-party plugin skill description length, for example 160-220 chars per skill.
• Add an aggregate budget check for the default enabled plugin set so the catalog does not exceed the metadata budget by itself.
• Keep description focused on trigger/routing signals; move detailed capability lists and workflow instructions into the body.
• Optionally split metadata into shortDescription / routing fields if long marketplace descriptions are needed elsewhere.

This would complement #19679: even if users can configure a larger budget, the shipped first-party catalog should stay compact by default.