meta-doctor

meta-doctor

Most monitoring tells you something broke. A doctor opens the pull request that fixes it.

meta-doctor is a Claude Code plugin marketplace that scaffolds autonomous "doctor" agents directly into your repository — scheduled health auditors that run on a recurring schedule, open PRs for code-fixable findings, file deduped tracker tickets for everything else, and post a digest after every run.

You describe your stack. The skill generates a production-ready audit-and-repair engine in your codebase, in your language — not a black-box SaaS, not a vendored binary you can't read.

Why this exists

Traditional observability stops at the alert. Someone still has to wake up, read the stack trace, find the file, write the patch, open the PR, and babysit CI. A doctor does that loop autonomously:

audit  →  classify  →  remediate (open PR)  or  file ticket  →  post digest

What a generated doctor actually does

Every scheduled run walks the same loop, and each stage is built to be trustworthy:

• Audits via runbooks. Each runbook is a Markdown audit definition (scan error logs, check data integrity, detect cost anomalies, …) executed through an agentic loop with real data-plane tools — log queries, metrics, SQL, source-control reads.
• Emits structured findings, not noise. Healthy runs emit zero findings. A finding is only raised for a confirmed, actionable problem — and every finding becomes a PR or a ticket, never a "just FYI" that piles up.
• Repairs code automatically — and conservatively. Code-fixable findings flow into a five-phase repair pipeline that localizes the bug (file → function → line), generates N independent patch candidates, and ships a PR only when a majority of candidates agree (self-consistency voting). This is the difference between "an LLM guessed a diff" and "three independent attempts converged on the same fix."
• Files everything else as deduped tickets. Non-code findings (compliance, data drift, cost) become tracker tickets, deduped against what's already open so you never get the same ticket twice.
• Babysits its own PRs. A repair-ci mode watches the doctor's open PRs, reads failing CI logs, and pushes fix commits — with a durable, restart-proof attempt counter so it never loops forever.
• Never goes dark. An error-isolation invariant guarantees the Slack/Teams/Discord digest posts at the end of every run, surfacing any remediation error so a human can step in. The expensive audit tokens are never wasted to a downstream failure.

Example: from log line to merged fix

A scan-logs run finds a TypeError firing 47 times in the last hour, traces it to a real source frame, and emits:

{
  "signature": "TypeError:scheduler/poller.ts:142",
  "severity": "high",
  "category": "code_bug",
  "title": "Uncaught TypeError in poller during backfill window",
  "description": "TypeError at poller.ts:142 when due_at falls inside a holiday exclusion window; the exception is swallowed and the send is silently dropped.",
  "suggestedFix": { "type": "code", "fileHint": "services/scheduler/src/poller.ts", "functionHint": "processWindow" }
}

The orchestrator dedupes it (no open PR for that signature yet), routes it to the repair pipeline, and you wake up to a PR titled "fix(scheduler): guard processWindow against undefined tz in holiday window" — already green, because repair-ci fixed the one lint error CI flagged overnight.

Install

In Claude Code, add this marketplace and install the plugin:

/plugin marketplace add michellemayes/meta-doctor
/plugin install meta-doctor@meta-doctor

Or from the CLI:

claude plugin marketplace add michellemayes/meta-doctor
claude plugin install meta-doctor@meta-doctor

What's included

This marketplace ships one plugin, meta-doctor, which provides two skills:

Once installed, the skills activate automatically when your request matches them (e.g. "scaffold a health-auditing agent for this repo" or "add a runbook that checks for stuck queue jobs").

Built for your stack, not a default one

The scaffolder isn't a template dump — it generates against your choices:

Budgets (max PRs/tickets per run, patch-candidate count, repair attempts, token caps) are all configurable and live in a single config.ts — no magic numbers buried in the engine.

Extending a doctor

Audit coverage is just Markdown. To add a check, run add-doctor-runbook and describe what to audit, what signals it reads, and what a violation looks like versus expected noise. The skill authors the runbook, assigns each finding a stable dedup signature, and registers it in the runbook index and schedule. Add a check today; it runs on the next scheduled cycle after redeploy.

Repository layout

.
├── .claude-plugin/
│   └── marketplace.json              # Marketplace catalog
└── plugins/
    └── meta-doctor/
        ├── .claude-plugin/
        │   └── plugin.json           # Plugin manifest
        └── skills/
            ├── scaffold-doctor/      # Generates the doctor engine
            │   ├── SKILL.md
            │   ├── references/       # Architecture, repair pipeline, finding
            │   │   └── …             #   contract, runbook authoring, swap table
            │   └── assets/runbooks/  # scan-logs example runbook
            └── add-doctor-runbook/   # Adds a runbook to an existing doctor
                └── SKILL.md

The references/ directory is the heart of the scaffolder — a set of stack-neutral contracts (module responsibilities, the five-phase repair pipeline, the finding schema, self-consistency winner selection) that Claude implements in your chosen language.

Validate

claude plugin validate .

License

MIT — see LICENSE.