mal-mcp: Skill for Claude Code

Docs/metadata consistency

Check every one of these, not just a sample:

• README.md's tool table matches src/tools/*.ts's registrations (names, and the auth-tier column — none / Client ID / OAuth token, per docs/auth.md — against what each tool actually needs).
• manifest.json's tools array lists the same tool names as what's actually registered (server.json has no tools array — the MCP Registry schema doesn't define one) — treat a test failure here as authoritative if one exists. Their description fields are deliberately short, independent marketing-style summaries, NOT a copy of the tool's full .describe()/description text in src/tools/*.ts — don't "fix" them to match verbatim, that's not a bug. Do re-read them for accuracy if a tool's behavior changed in a way the short summary now misrepresents.
• Tool description/field .describe() text in src/tools/*.ts itself: does it still match the actual inputSchema/outputSchema and the real behavior? Cross-check new/edited descriptions against the tool-description-check skill (Glama's TDQS rubric) per AGENTS.md.
• CHANGELOG.md's [Unreleased] section (see the changelog-style skill for entry style) has one line per real behavior change made in this pass — add missing entries, don't just flag them as missing. Run node scripts/check-changelog-coverage.mjs to list every commit since the last release tag and flag which ones CHANGELOG.md doesn't reference at all — it's not a hard gate (plenty of commits are legitimately internal: dev-dependency bumps, docs-about-docs, CI-only fixes), but triage every line it reports as either "needs an entry" or "correctly excluded," don't skip the check just because most commits usually are internal.
• docs/api-references.md's "verified live" claims still match the current client code, especially any claim this pass's own fixes just invalidated — and especially the official-API fallback field-gap list, which is exactly the kind of claim a MAL API change could quietly break.
• docs/auth.md's credential-tier breakdown still matches what each tool actually requires.
• AGENTS.md's project-shape/file-tree description (including this .agents/skills/ entry) still matches the filesystem.
• notes/tenrai-reliability.md (gitignored) — if this pass turned up a new Tenrai quirk or reliability data point, log it there with a date, the same way past passes have; don't let a fresh finding live only in this conversation's transcript.
• PRIVACY.md and SECURITY.md: re-verify every specific claim against the actual current code, don't just skim for plausibility — which credentials exist and how each is transmitted/redacted (e.g. src/lib/errors.ts's redact() actually covering both key=value and JSON "key":"value" shapes, and the tokens.json store's 0600/0700 permissions), what is and isn't cached (incl. cache key/TTL — cross-check the "deliberately not cached" list against src/lib/cache.ts call sites), the current list of read-only vs. OAuth/write tools, and the host-allowlist statement (config.ts's Zod validation vs. an actual fixed allowlist — there isn't one). This class of drift is easy to miss because it reads fine on its own and only breaks against the code: a sibling repo's SECURITY.md/PRIVACY.md both claimed "player-specific data is never cached" after a later feature added exactly that caching, and a separate claim conflated an actually-cached field with a similarly-named never-cached one — neither doc was self-evidently wrong, both required re-reading the client code to catch. Also confirm every root-level doc a README link points to (PRIVACY.md, SECURITY.md, …) is actually in package.json's files array — a new one added there without updating files 404s on the npm tarball (bit PRIVACY.md once already, then recurred for SECURITY.md).