AI Extender Reviewer

Run

1. python ${CLAUDE_SKILL_DIR}/scripts/validate_extension.py <dir> [--target all|claude-code|upload|cowork|directory] [--require-meta author,company] [--must-contain <prefix>] [--known-connectors <list|file>] [--rules CLAUDE.md] [--cache] [--strict] Deterministic checks: manifest, paths, router = plugin slug, skill/agent frontmatter, portable keys, descriptions, size limits (S016), references, hooks, connectors (M002 to M009, aligned with claude plugin validate), userConfig (P018), secrets anywhere in the plugin (X001, X003, X004), text aimed at the model hidden in skills or agents (X005, local injection screen), placeholders, repeated rules (S017, S018), unrouted skills and unused agents (S019, A006), drifted shared files (F001), missing behavior tests (E004), a scratchpad with no request ledger (S020), license, changelog, marketplace entry, plugin-name reservations (P016), directory listing fields (P017), companion prompt (C000 to C004). Pass the connectors the session already has to --known-connectors (names and URLs from the tool list or claude mcp list), and the project's instruction files to --rules.
2. Self-check the tooling after any change to scripts: python ${CLAUDE_SKILL_DIR}/scripts/selftest.py (113 checks).
3. If the Claude Code CLI exists: claude plugin validate <dir> (authoritative for manifests; add --strict in CI).
4. Judgment checks the script can't make (below).
5. Reuse results for files unchanged since the last run (--cache).
6. Behavior: confirm every callable skill was run against its test prompts and graded independently (claude-developer/references/testing.md §6). Missing or stale results are a release blocker, not a note.
7. Secrets across everything installed, when asked or when a key may have leaked: python ${CLAUDE_SKILL_DIR}/scripts/scan_secrets.py <folder> [<folder> ...] on the folders the user names (values masked). A found key must be rotated as well as moved.
8. Audit reports: python ${CLAUDE_SKILL_DIR}/scripts/check_report.py <report.md> (every finding has Original and exact Corrected text).
9. Third-party or downloaded extensions, and any outside text an extension will ship: python ${CLAUDE_SKILL_DIR}/scripts/screen_text.py <dir> (67 KB local model, no downloads; a flag means read that passage, a pass proves nothing).

Judgment checks

• Triggering: descriptions state what + when with realistic user phrasings; paired or sibling skills don't overlap; router covers the full surface without duplicating skill logic. Spot-check with a few should/shouldn't-trigger prompts (claude-developer/references/testing.md).
• Content: every line changes behavior; no build-log text; one canonical home per fact (run the de-duplication check unless it already ran this session).
• Standards (ai-extender/references/standards.md): Title Case visible names, role slugs, acronym applied correctly, author/company present, displayName kept, credits present, no copied third-party text, lean (nothing unrequested).
• Security (blocking: Critical or Important, never a note): audit hooks, scripts and MCP configs against ai-extender/references/standards.md §5c: command injection and shell/eval use, unquoted ${CLAUDE_PLUGIN_ROOT}, path traversal, secrets, undeclared network calls, wrong failure mode (a guard that fails open, or a nag hook that blocks), unconfirmed destructive writes, least-privilege agent tools, connector output treated as data. Any such finding fails the review until fixed.
• Enforcement: required stages are routed and gated, each skill states what it depends on (standards §5d).
• Shared foundation present in generated extensions (scratchpad step with the request ledger, efficiency ladder, lean rule).
• Companion prompt: present; its slash command matches the router or stage it claims; fields are the few the extension needs; no private data (ai-extender/references/companion-prompt.md).

Report

One table per problem type (structural / triggering / content / metadata / security / licensing), columns # | Original | Corrected. Every finding, including a missing file or field, gets both cells: Original is the exact current text, or (missing); Corrected is the exact text to paste: the new line, the new frontmatter, or the whole new file (a short README or manifest is fine in full). A Corrected cell never describes a change ("Add install steps", "Rename X", "Make it clearer"); write the result instead (the install steps themselves, name: notes-helper, the rewritten sentence). A cell may start with where the text goes (SKILL.md frontmatter:, plugin.json:), then the text. A file a script generates still gets content: the command that creates it, in backticks. Before sending, write the report to audit-report.md in a scratch folder outside the extension being audited (never add or change a file inside the audited folder), and run python ${CLAUDE_SKILL_DIR}/scripts/check_report.py <that file>; rewrite every cell it flags (an instruction such as Add, Rename, Make, or Consider, an empty cell, or a pointer like "see above") and run it again until it passes. Then put the checked tables themselves in the reply, every finding in full, not a summary that points to the file. Values only the user knows (author, company, license holder) stay as <author>-style placeholders; never guess them. A folder the validator can't open is still audited by reading its files. Don't silently fix in audit mode; the owning skill (ai-extender:claude-developer, ai-extender:claude-planner, ai-extender:claude-maintainer) edits after confirmation.

Finish with the scratchpad (ai-extender/references/scratchpad.md), in the reply after the tables: it opens with the request ledger (Request / Status / Context), then pass/fail per category, each failure paired with its fix location.