xpd
Run uvx xpd check . to find the lines of your Python code that are slow or give a wrong
result, and add --fix to let xpd rewrite the ones it can. xpd checks code that uses pandas,
NumPy, SciPy, scikit-learn, scikit-lego, MAPIE, Category Encoders, linearmodels, arch, pymoo,
plotly or Apache Hamilton. It starts no Python, so a coding agent can run it after every edit.
$ uvx xpd check samples/bad.py
samples/bad.py:5:15: warning [pandas/row-loop] iterrows() builds a Series for every row. Compute on whole columns instead, such as df["a"] * df["w"]. When each row has to become a Python object, loop over itertuples() or to_dict("records").
samples/bad.py:6:5: error [pandas/chained-assignment] Chained assignment writes to a copy under copy-on-write, so the frame keeps its old values. Assign to df.loc[i, "a"] instead.
samples/bad.py:7:1: info [pandas/no-inplace] inplace=True gives no speedup and blocks chaining. Reassign instead.
samples/bad.py:8:5: warning [pandas/merge-without-validate] merge without validate= can silently duplicate rows.
samples/bad.py:11:5: warning [pandas/concat-in-loop] out grows by one concat per pass, and each concat copies every earlier row. Collect the frames in a list and concat once after the loop.
xpd: 5 issues, 2 more with --fix --unsafe-fixes
What a fix changes
xpd check samples/bad.py --fix --unsafe-fixes rewrites two of the five lines and reports the
other three again.
import pandas as pd
df = pd.DataFrame({"a": [1, 2], "k": [1, 2]})
other = df.copy()
for i, row in df.iterrows():
- df["a"][i] = row["a"] * 2
-df.dropna(inplace=True)
+ df.loc[i, "a"] = row["a"] * 2
+df = df.dropna()
m = pd.merge(df, other, on="k")
out = pd.DataFrame()
for f in ["x.csv"]:
out = pd.concat([out, pd.read_csv(f)])
Fix every error, and fix a warning unless you can say why the code is right in that place.
| Word | Meaning |
|---|---|
error | The code gives a wrong result or raises. |
warning | The code is slow or easy to get wrong. |
info | A shorter or newer form exists. |
| safe fix | The rewrite keeps what the code does, and --fix writes it. |
| unsafe fix | The rewrite may change what the code does. --fix --unsafe-fixes writes it, so read the diff afterwards. |
xpd check exits with 0 when it finds no error, and with 1 when it finds one. It exits with 2
when it can't run.
Install
uv tool install xpd installs the xpd command, and uvx xpd check . runs it once without an
install.
uv tool install xpd
Rules
Each library has one pack of rules, and the pack runs only on a file that imports the library. The Python pack runs on every file, and the plotly pack also runs on a file that makes plotly the plotting backend of pandas.
| Pack | One of its rules | What the rule reports |
|---|---|---|
| pandas | chained-assignment | df["a"][i] = v writes to a copy, so the frame keeps its old values |
| NumPy | append-flattens | np.append without axis flattens both arrays |
| SciPy | removed-keyword | tol= on the sparse iterative solvers was removed and raises TypeError |
| scikit-learn | fit-on-test | A fit on the test rows scored 0.997 AUC on data with no signal |
| scikit-lego | thresholder-score | Thresholder.score ignores threshold, so a grid search over 0.1, 0.5 and 0.9 gave 0.887 three times |
| MAPIE | calibrate-on-training-rows | conformalize on the training rows covered 0.468 of the test rows where 0.9 was asked |
| Category Encoders | ordinal-mapping-pairs | An OrdinalEncoder mapping given as a list of pairs codes every row as -1.0 |
| linearmodels | clustered-without-clusters | cov_type="clustered" without clusters gives the robust standard errors |
| arch | unscaled-returns | Returns as decimals stopped the optimizer with code 4 and gave a mean of 4.02 where 0.056 was right |
| pymoo | objectives-stacked-by-row | out["F"] = np.array([f1, f2]) in a Problem gives each solution the wrong values |
| plotly | marker-size-diameter | A marker's size is its diameter, so markers of size 10 and 20 had areas of 78.5 and 314.2 square pixels |
| Hamilton | node-mutates-input | A node writes to its input, and Hamilton hands the same object to other nodes |
| Python | dict-str-any | dict[str, Any] in a signature hides the keys the function uses |
xpd rules lists every rule, and the rules page
shows each one with code before and after.
Commands
xpd <command> -h lists the flags of a command, and the
commands page shows the output and the exit
codes of each one.
xpd
├── lint
│ ├── check [paths] lint files and folders, and rewrite with --fix (alias: scan)
│ ├── rules list every rule with its documentation
│ └── hook lint the files that an agent's edit changed
├── look up a library
│ ├── find "<task>" find library symbols, rules and worked recipes for a task
│ ├── api <name> read one symbol, method or recipe in full
│ └── index <source> build a search index from library source
├── read and rewrite code
│ ├── run [paths] find a code pattern, and preview or apply a rewrite (alias: query)
│ ├── outline [paths] show imports, definitions and members with line numbers
│ └── graph <files> print the dataflow of step code or Hamilton modules
├── run the project's Python
│ ├── profile <data file> print dtypes, nulls, unique counts and memory per column
│ ├── equiv old.py new.py run two scripts and compare the `result` each one leaves
│ └── refit <contract> refit a saved Pipeline on raw training rows and compare
├── write your own rules
│ ├── new create a rule project, a rule, a test case or a utility
│ └── test check rule examples and compare snapshots
└── connect a tool
├── mcp start the MCP server for agents
├── lsp start the language server for editors
├── skill print the agent skill for the installed version
└── completions <shell> print shell completion for bash, elvish, fish, powershell or zsh
profile, equiv and refit run the project's Python, and they need pandas in the project's
environment. index runs it too, with Griffe, and the first find runs it once to download the
search model. mcp starts the MCP server with the Python that the xpd package installs. Every
other command only reads files.
Look up a library you don't know
xpd find "<task>" lists the classes, rules and recipes for a task in your own words, even for
a library you haven't installed. A recipe is a worked example from the library's documentation.
xpd api <name> then prints one entry in full, with each parameter's type annotation and default
as the source writes them, so a default in a stub file shows as .... You can also name a method
of a class, as the last call below does with mapie.subsample.BlockBootstrap.split.
$ xpd find "time series bootstrap for cross validation" --limit 1
Symbols
sklearn.model_selection.TimeSeriesSplit(n_splits, max_train_size, test_size, gap) [class]
Time Series cross-validator.
Rules: sklearn/groups-ignored, sklearn/removed-in-sklearn
Rules
sklego/group-time-series-split-folds [rule]
The rule reports `GroupTimeSeriesSplit` with 6 or more splits. The class tries every combination of
split points over the groups to find the most even folds, so each extra fold multiplies the time.
Recipes
mapie/recipe/generate-a-sampling-method-that-block-bootstraps-the [recipe]
Generate a sampling method, that block bootstraps the training set.
$ xpd api mapie.subsample.BlockBootstrap
mapie.subsample.BlockBootstrap(n_resamplings: int = 30, length: Optional[int] = None, n_blocks: Optional[int] = None, overlapping: bool = False, random_state: Optional[Union[int, RandomState]] = None) [class]
Generate a sampling method, that block bootstraps the training set. It can replace KFold, LeaveOneOut or SubSample as cv argument in the TimeSeriesRegressor class.
get_n_splits(args, kargs) Returns the number of splitting iterations in the cross-validator.
split(X, args, kargs) Generate indices to split data into training and test sets.
Rules: mapie/unseeded-resampling
$ xpd api mapie.subsample.BlockBootstrap.split
mapie.subsample.BlockBootstrap.split(X: NDArray, *args: Any, **kargs: Any) [method]
Generate indices to split data into training and test sets.
Fact about find | Value |
|---|---|
Results without --limit | Up to 5 symbols, 3 rules and 3 recipes |
| Flags | --kind symbol, --kind rule or --kind recipe selects one kind, and --package selects one library |
| Ranking | By keywords and by meaning, so a task finds an entry whose documentation uses other words |
| Model for the meaning | EmbeddingGemma, a 333 MB file that the first find downloads once |
| Without the model | find says so on stderr and ranks by keywords only |
| Where recipes come from | Library documentation and galleries. xpd keeps an example only when its own rules report nothing in it, and it never runs the example |
The first find saves the model in the folder that the table below names for your system, and
XDG_CACHE_HOME takes precedence on every system.
| System | Model folder |
|---|---|
Any, with XDG_CACHE_HOME set | $XDG_CACHE_HOME/xpd/models |
| Linux | ~/.cache/xpd/models |
| macOS | ~/Library/Caches/xpd/models |
| Windows | %LOCALAPPDATA%\xpd\models |
Use in an editor
xpd lsp starts a language server, so your editor shows each issue while you type, with a quick
fix and the rule's documentation on hover. The
language server page has the setup for Neovim, Helix,
Zed and VS Code.
Use with a coding agent
Pick the plugin when your client supports it, because it sets up the MCP server and the skills together. Otherwise set up the parts you want one at a time.
| Way | What the agent gets | Setup |
|---|---|---|
| Plugin | The MCP server and the skills together | Load the xpd repository in a client that supports Agent Plugins, such as Codex |
| Hook | The issues of each file, right after the edit | xpd hook in the Claude Code settings |
| MCP server | Typed tools, the rules, the skills and prompts | xpd mcp |
| Skills | When to run each command, and Python style that a linter cannot check | npx skills add bjoernaagaard/xpd |
Plugin
Run the commands below to install the plugin in Codex. The plugin starts xpd with uvx, so the
machine needs uv.
codex plugin marketplace add bjoernaagaard/xpd
codex plugin add xpd@xpd
The repository is a plugin in the Agent Plugins format, version 1.0.0. Version 1.0.0 has no hooks, so you add the hook below yourself.
xpd, loaded as a plugin
├── plugin.json the name and version of the plugin
├── skills/ xpd, python-best-practices and python-review
├── mcp.json starts the MCP server of the same xpd version with uvx
└── .agents/plugins/marketplace.json lists the plugin for Codex
Hook
Add the hook below to .claude/settings.json in your project, and Claude Code shows the agent
the issues of each .py file it edits, in the same turn.
{
"hooks": {
"PostToolUse": [
{
"matcher": "Edit|Write",
"hooks": [{ "type": "command", "command": "xpd hook" }]
}
]
}
}
Claude Code edits features.py
└── the PostToolUse hook runs xpd hook
├── no issues: exit code 0
├── issues: exit code 2, and Claude Code shows them to the agent
└── xpd can't run: exit code 1, and Claude Code shows the message to you
MCP server
The command below adds the MCP server to Claude Code, so the agent calls xpd through typed tools and needs no shell.
claude mcp add xpd -- uvx xpd mcp
xpd mcp
├── tools
│ ├── check, fix lint, and write fixes after the user confirms unsafe ones
│ ├── graph, profile, equiv, refit the commands of the same names
│ ├── find, api look up a library for a task and read its API
│ ├── query, rewrite search code, preview replacements and confirm edits
│ ├── outline read the project's symbols and imports
│ ├── new, test create rules, and check their examples and snapshots
│ ├── rules, rule list the rules, and read one with its examples
│ ├── discovery_gate, contract, gates judge a run from a data file to a saved model
│ └── show_issues, show_dataflow, show_columns tables and a chart, in hosts that draw them
├── resources
│ ├── xpd://rules/{package}/{rule} one rule's documentation
│ └── skill://<name>/SKILL.md each skill that ships with xpd
└── prompts
└── lint_and_fix, review, refactor_data_logic, add_to_dataflow, build_pipeline
The server takes every rule and every result from the xpd binary, so it matches the installed version. The MCP page covers other clients, the settings and how to serve xpd over HTTP. The code mode page shows a run from a data file to a saved scikit-learn model, with checks that xpd runs itself.
Skills
The command below installs the three Agent Skills in skills/. The
two Python skills work without xpd, and they run xpd when it is installed.
npx skills add bjoernaagaard/xpd
| Skill | What it does |
|---|---|
xpd | Says when to run each xpd command. xpd skill prints it for the installed version. |
python-best-practices | Holds the rules an agent follows when it writes or changes Python. |
python-review | Reviews a file, package or diff, and cites the rule for every finding. |
Documentation
The full guide is at https://bjoernaagaard.github.io/xpd/, and its rules page also explains how to write a rule.
License
You may use xpd's own code under the MIT license or the Apache-2.0 license, and the other parts that ship with xpd keep the licenses in the table.
| Part | License | Where the text is |
|---|---|---|
| xpd's own code | MIT or Apache-2.0, as you choose | LICENSE-MIT and LICENSE-APACHE |
| Recipes, which are examples taken from library documentation | The license of the library, which is MIT, BSD-3-Clause, Apache-2.0 or NCSA | licenses/, one file per library. xpd api <recipe> prints the license name and the source link with the code |
option-ext, a dependency shipped unmodified | MPL-2.0 | licenses/option-ext-LICENSE.txt, with the complete source in licenses/option-ext-0.2.0-source.tar.gz and at https://crates.io/api/v1/crates/option-ext/0.2.0/download |
libcst, the parser that checks a fixed file | MIT, with parts of CPython under PSF-2.0 | licenses/libcst-LICENSE |
EmbeddingGemma, the model that find downloads | The Gemma Terms of Use at https://ai.google.dev/gemma/terms | Not shipped with xpd |
The licenses/ folder ships in the crate, the sdist and the wheel, and the package metadata
lists the combined licenses.