confluent-skill-reviewer — audit a Confluent agent skill

Three bars every skill in this repo must clear:

1. Agent Skills spec — frontmatter shape, naming rules, directory layout, token budgets (https://agentskills.io/specification).
2. Confluent conventions — lazy-loaded references, anti-trigger clauses, mode-table branching, ≥90% evals, SME + DTX/DevRel sign-off (CLAUDE.md, .github/pull_request_template.md).
3. Evals-as-contract — evals/evals.json with specific, verifiable expectations; fixtures kept in sync.

This skill walks an agent through a structured audit and returns a single Markdown report. It is read-only by default — never modifies the skill under review.

Pre-flight gates

Before producing any findings, confirm all three out loud (briefly):

1. Scope: which mode are you in? See Mode Detection. If unclear, ask the user once.
2. Target paths exist: list the skills/<name>/ directories you will audit. Stop if none.
3. Tool availability: run bash skills/confluent-skill-reviewer/scripts/run_skill_validator.sh --probe (from the repo root) to check whether the external skill-validator binary is installed. If absent, note it in the report and continue with native checks — do not block.

Skipping these gates is the most common source of bad reviews. The point is to be explicit about scope so the user can redirect early.