Add /ai-coding-config doctor — diagnose plugin health, hook state, and config drift

[TechNickAI]

Problem

We have `/ai-coding-config` for setup and updates, but no diagnostic mode. Users (and Nick) hit issues like:

• Plugin marketplace registered but not actually installed.
• Symlinks broken (`.claude/commands` → `plugins/core/commands` etc.) after a move or clone.
• `settings.json` / `settings.local.json` drift between repo and user install.
• Hooks declared but not executable (missing shebang, wrong permissions).
• Version skew: user's local `ai-coding-config.md` command is older than the repo's.
• MCP servers referenced by skills not actually installed.

There's no single command to answer "is my install healthy?" Each issue requires manual investigation.

Reference

oh-my-claudecode ships an `omc doctor` command. Their pattern: run a battery of checks, print a checklist with pass/fail/warn, suggest fixes. Not novel as a concept, but worth borrowing the shape.

The context-mode plugin we use also exposes `/ctx-doctor` — same pattern, narrower scope. Worth looking at how it structures its output.

What to build

A new slash command `/ai-coding-config doctor` (or `/ai-coding-config-doctor` if naming requires it) that runs a diagnostic battery and reports as a checklist.

Suggested checks (research and refine)

The implementer should walk the repo and figure out what actually breaks in practice, not just what theoretically could. Starting points:

Plugin install state

• Marketplace registered? (`claude plugin list` or equivalent)
• All plugins in `.claude-plugin/marketplace.json` actually installed?

Symlinks

• `.claude/commands`, `.claude/agents`, `.claude/skills` resolve to `plugins/core/*`?
• `rules/` → `.cursor/rules/` intact?

File integrity

• All commands referenced in marketplace.json exist on disk?
• All agents referenced exist? Skills?
• Hook scripts executable and have valid shebangs?

Settings

• `settings.json` parses as valid JSON?
• `settings.local.json` doesn't override critical hooks unintentionally?
• Permissions list matches what plugins declare they need?

Version drift

• Repo VERSION newer than installed? (mirror the existing `/ai-coding-config update` check)
• Local copies of commands (e.g. `.claude/commands/ai-coding-config.md` per the existing update protocol) outdated?

MCP servers

• Servers referenced in skills (context7, sentry, slack, etc.) actually connectable?
• Auth set up for the ones that need it?

Skills

• Frontmatter parses?
• `triggers` field present where expected?
• (After feat: Transform into Claude Code plugin marketplace #2 lands) new pipeline frontmatter validates?

Output format

Mirror `/ctx-doctor` style: ✅ / ⚠️ / ❌ checklist with one-line explanations. Group by category. End with a "suggested fixes" section that links each failure to the command that resolves it (often `/ai-coding-config update` or a manual step).

Open questions

• Should it offer to auto-fix safe issues (broken symlinks, file permissions) or always be read-only?
• Should it work in any project that uses ai-coding-config, or only in the source repo?
• Does it need a JSON output mode for CI / scripting?

Success criteria

• Running `/ai-coding-config doctor` in a fresh install gives all green.
• Running it after a known broken state (e.g. `rm .claude/commands`) surfaces the breakage with a clear fix.
• The check list is comprehensive enough that "my install is broken" stops being a thing Nick has to debug manually.

[TechNickAI]

Taking this on. Approach:

• Extend the existing /ai-coding-config command with a doctor subcommand (same file, new section — users invoke as /ai-coding-config doctor)
• Checks: symlink integrity, hook executability/shebang, settings.json validity, hooks registered, version drift, frontmatter parse, marketplace JSON validity
• Output: ✅/⚠️/❌ checklist grouped by category, with fix suggestions at the end

[TechNickAI]

PR #56 implements the /ai-coding-config doctor subcommand: #56

The doctor runs a ✅/⚠️/❌ checklist grouped by category: plugin install state, symlinks (source repo), hook scripts, marketplace JSON consistency, version drift, and skill frontmatter validation. It's context-aware (source repo vs user project), read-only except for hook permission repair, and prints results progressively.