aidoctor 0.1.1

Your agent writes bad Python. This catches it.

A static analyzer for AI-generated Python code. Catches 25 patterns Claude Code, Cursor, Copilot, Codex, Gemini, and friends produce: bare except, hardcoded secrets, async/sync mismatch, dead defenses, fake type hints, stub comments shipped as production code. Then installs a skill into your agent so it stops writing them.

Install · CLI · Rules · Leaderboard · vs alternatives

Install

Install differs by harness. If you use more than one, install aidoctor separately for each.

Claude Code

Register the marketplace and install the plugin:

/plugin marketplace add ankit-aglawe/aidoctor
/plugin install aidoctor@ankit-aglawe

After install, just talk to Claude in plain English. Three skills auto-load:

• scan — runs aidoctor, summarizes the top findings, asks before fixing
• simplify — three-angle review (reuse / quality / efficiency), then fixes
• python-rules — auto-loads when you write Python; your agent reads the 25 rules first

Slash forms /aidoctor:scan and /aidoctor:simplify work too.

Cursor

In Cursor Agent chat:

/add-plugin aidoctor

Or run uvx aidoctor install once — it writes ~/.cursor/rules/aidoctor.mdc.

Codex CLI

Drop the rules file into Codex's config dir:

uvx aidoctor install

Writes ~/.codex/skills/aidoctor.md. Codex reads it automatically when generating Python.

Gemini CLI

uvx aidoctor install

Writes ~/.gemini/skills/aidoctor.md.

OpenCode

uvx aidoctor install

Writes ~/.config/opencode/rules/aidoctor.md.

Any other agent (Aider, Copilot, custom)

uvx aidoctor skill --format generic > my-agent/rules/aidoctor.md

Or paste this prompt into the agent:

Install aidoctor and run my first scan.

1. Run uvx aidoctor install (or pip install aidoctor && aidoctor install if uvx is missing).
2. Run uvx aidoctor scan . in the current project.
3. Summarize in one paragraph: (a) the score out of 100 and its label, (b) the top 3 rule violations and what each means in plain English, (c) whether to fix errors or warnings first.
4. Do not modify any files. Stop after the summary.

CLI

For humans and CI:

uvx aidoctor scan .             # zero-install
uv tool install aidoctor        # persistent (2026-native)
pipx install aidoctor           # isolated
pip install aidoctor            # classic / CI

Then:

aidoctor scan .                       # full repo
aidoctor scan src/ tests/             # specific paths
aidoctor scan --diff                  # only lines you've changed
aidoctor scan-pr https://github.com/owner/repo/pull/42
aidoctor scan --explain bare-except-pass    # docs for one rule

No signup. No API key. No telemetry. Runs entirely on your machine.

What it catches

25 rules across 8 categories. Each rule has a stable ID (bare-except-pass, hardcoded-api-key, range-len-loop) that appears identically in scan output, the skill markdown, and the slash command — so an agent can cite a finding back to you and you can cite a finding back to the agent.

Category	Rules
Dead defenses	bare except: pass, except Exception swallowing, unreachable raise after return, redundant null-check after isinstance
Hardcoded secrets	API key / token literals, AWS credentials, JWT-shaped strings
Async/sync mismatch	sync I/O in async fn, asyncio.run inside async fn, blocking call in event loop
Fake type hints	Any everywhere, missing return type on public fn, generic without TypeVar
Stale loop patterns	mutate list during iteration, range(len(x)), time.sleep in tests
Performance	nested-loop append, += string concat in loop, repeated dict lookup
AI-slop imports	wildcard import, duplicate import, conditional import outside try, import without use
Comment-driven decay	TODO/FIXME without ticket, stub comments (# implement this) shipped as code

Full rule reference: aidoctor scan --explain <rule-id>.

Score

aidoctor scan outputs a 0–100 health score. The score penalizes the number of unique rules tripped, not raw violation count — so fixing one category moves the number, instead of chasing per-line totals.

Band	Label
90–100	Healthy
70–89	Needs work
0–69	Critical

Same formula on every machine. Same formula in CI.

Slash commands

In Claude Code, the plugin installs /aidoctor:scan (run a scan + summary). In OpenCode and Gemini CLI, aidoctor install drops a global /aidoctor command. Cursor and Codex don't support custom slash commands; the rules file is the install vector there.

GitHub Action

- uses: ankit-aglawe/aidoctor-action@v1
  with:
    fail-on: error      # error | warning | none

Or call the CLI directly in any workflow:

- run: uvx aidoctor scan . --fail-on error

Pre-commit:

repos:
  - repo: https://github.com/ankit-aglawe/aidoctor
    rev: v0.1.0
    hooks:
      - id: aidoctor

Configuration

aidoctor reads [tool.aidoctor] in pyproject.toml if present. The defaults are designed to be useful without any config.

[tool.aidoctor]
exclude = ["migrations/", "vendor/"]
fail-on = "error"

Inline suppression

# aidoctor: disable=hardcoded-api-key
API_KEY = "sk-test-not-real"

Variants:

foo()  # aidoctor: disable-line=range-len-loop

# aidoctor: disable-file=stub-comment,todo-without-ticket

Multiple rules: # aidoctor: disable=rule-1,rule-2.

How it differs

	aidoctor	Ruff	Sloppylint	CodeRabbit
Catches AI-author patterns specifically	✓	partial	✓	✓
Stable rule IDs	✓	✓	—	per-customer
Installs a skill in your agent	✓	—	—	—
Runs locally, no cloud	✓	✓	✓	—
Free CLI	✓	✓	✓	$24/seat/mo
Per-PR scan	✓ (scan-pr)	via Action	—	✓

Ruff is the right tool for correctness. aidoctor is the right tool for the specific failure modes LLMs have when they write Python — patterns that pass Ruff and mypy on default settings but break in production.

Leaderboard

How major Python projects score:

Repo	Score	Top issues
coming at launch	—	—

Want your project listed? Open a PR adding it to leaderboard.yaml.

Roadmap

• Claude Code plugin via /plugin marketplace add ankit-aglawe/aidoctor
• Listed on the official Anthropic plugin marketplace
• MCP server (aidoctor mcp) so Cursor / Windsurf / Codex / Gemini reach the rules over a single transport
• aidoctor learn — propose project-local rules from your git history
• PR-delta scoring on GitHub Action

Credits

Inspired by react-doctor by @aidenybai. Built for the era where most Python isn't written by humans anymore.

License

MIT