scitex-io 0.2.20

Universal scientific data I/O with plugin registry

SciTeX IO (scitex-io)

Universal scientific data I/O with plugin registry

Full Documentation · uv pip install scitex-io[all]

---

Problem and Solution

#	Problem	Solution
1	Format zoo — every format has its own API (pd.read_csv, np.load, pickle, h5py, torch.save, …).	One call — sio.save(obj, "x.ext") / sio.load("x.ext") dispatches across 30+ formats.
2	Outputs scatter — saves land relative to cwd, drift away from the script that produced them.	Caller-anchored paths — sio.save(df, "results.csv") writes to {caller}_out/results.csv.
3	Magic numbers everywhere — hyperparameters, paths, thresholds duplicated across scripts.	Centralized config — load_configs() merges every config/*.yaml into one DotDict.

Quick Start

import scitex_io as sio
import pandas as pd
import numpy as np

# Demo Data
df_orig = pd.DataFrame({"x": [1, 2, 3]})
arr_orig = np.array([1, 2, 3])
params_orig = {"lr": 1e-3, "epochs": 10}

# Unified Saving API
sio.save(df_orig, "data.csv")
sio.save(arr_orig, "data.npy")
sio.save(params_orig, "config.yaml")

# Unified Loading API
df_loaded = sio.load("data.csv")
arr_loaded = sio.load("data.npy")
params_loaded = sio.load("config.yaml")

# Round-trip check
assert df_loaded.equals(df_orig)
assert np.array_equal(arr_loaded, arr_orig)
assert params_loaded == params_orig

Supported Formats (30+) and Customization

Category	Extensions
Spreadsheet	.csv, .tsv, .xlsx, .xls, .xlsm, .xlsb
Columnar	.parquet, .feather
Scientific	.npy, .npz, .mat, .hdf5, .h5, .zarr
Serialization	.pkl, .pickle, .pkl.gz, .joblib
ML/DL	.pth, .pt, .cbm
Config	.json, .yaml, .yml, .xml
Database	.db (SQLite3)
Documents	.txt, .md, .pdf, .docx, .tex, .log
Code	.py, .sh, .css, .js
Images	.png, .jpg, .jpeg, .gif, .tiff, .tif, .svg
Media	.mp4
Web	.html
Bibliography	.bib
EEG	.vhdr, .vmrk, .edf, .bdf, .gdf, .cnt, .egi, .eeg, .set, .con

Need a format not listed above? Register a custom handler with register_saver / register_loader and sio.save() / sio.load() will dispatch to it by extension just like a built-in.

from scitex_io import register_saver, register_loader

@register_saver(".custom")
def save_custom(obj, path, **kw):
    open(path, "w").write(str(obj))

@register_loader(".custom")
def load_custom(path, **kw):
    return open(path).read()

sio.save("hello", "data.custom")
assert sio.load("data.custom") == "hello"

Installation

uv pip install "scitex-io[all]"

Per-module extras

Extra	Pulls in
scientific	scipy, h5py, zarr, numcodecs, matplotlib (HDF5 / zarr / scientific I/O)
mcp	fastmcp (MCP server for agents)
all	scientific + mcp (recommended)
dev	pytest, pytest-cov, plotly, Pillow, + every optional dep so the test suite runs
docs	Sphinx + RTD theme + myst-parser (docs build only)

uv pip install "scitex-io[scientific]"   # HDF5 / zarr / parquet stack
uv pip install "scitex-io[mcp]"          # MCP server only
uv pip install -e ".[dev]"               # editable install for contributors

How it works

1. Format detection by extension

save() / load() pick the right reader/writer from the file extension via a plugin registry — just as OS does. Custom handlers can be available using register_saver / register_loader.

%%{init: {'flowchart': {'nodeSpacing': 20, 'rankSpacing': 40, 'curve': 'linear'}, 'themeVariables': {'fontSize': '12px'}}}%%
flowchart LR
    A["save(obj, 'x.ext')"] --> B{Registry}
    L["load('x.ext')"] --> B
    B -->|.csv .parquet .feather| C[pandas]
    B -->|.npy .npz| D[numpy]
    B -->|.h5 .zarr| E[h5py / zarr]
    B -->|.pkl .joblib| F[pickle / joblib]
    B -->|.pt .pth| G[torch]
    B -->|.png .jpg .svg| H[Pillow]
    B -->|.yaml .json| I[PyYAML / json]
    B -->|.bib .pdf .docx ...| J[30+ handlers]
    B -.->|register_*| K[Custom format]

2. save(obj, out.ext) in /path/to/script.py → /path/to/script_out/out.ext

Relative paths in save() resolve relative to the calling script / notebook, not the working directory. Scripts and outputs are tied as locations.

/path/to/project/
├── config/                              # see §3
│   └── ...
└── scripts/
    └── xxx/
        ├── filename.py                  # sio.save(df, "results.csv")
        └── filename_out/                # auto-created sibling of the script
            └── results.csv              # output lands here

Caller	sio.save(df, "sub/dir/results.csv") writes to
/path/to/analysis.py (script)	/path/to/analysis_out/sub/dir/results.csv
/path/to/exp.ipynb (notebook)	/path/to/exp_out/sub/dir/results.csv
python -i / IPython / REPL	~/.scitex/io/runtime/cache/sub/dir/results.csv

Bare filename or any relative path — "results.csv", "sub/dir/results.csv", and "./sub/dir/results.csv" all work; the whole path is appended under the caller's output anchor.

Intermediate directories created automatically — no os.makedirs() / Path.mkdir() calls needed on the caller side.

Advanced save() — absolute paths, symlinks, dry-run

Absolute paths bypass auto-routing. sio.save(df, "/data/x.csv") writes to /data/x.csv as-is — caller-anchored routing (§2) only applies when the path is relative.

sio.save(df, "/data/x.csv")                            # absolute → used as-is

Symlinks and dry-run. symlink_from_cwd=True drops a symlink at ./results.csv pointing into the auto-routed location; symlink_to=… plants a symlink at a custom path; dry_run=True prints the resolved path without writing.

sio.save(df,  "results.csv", symlink_from_cwd=True)
sio.save(fig, "fig1.png",    symlink_to="/data/latest/fig1.png")
sio.save(df,  "results.csv", use_caller_path=True)     # resolve from caller script
sio.save(df,  "results.csv", dry_run=True)             # print path, don't write

3. Centralized project configuration

Scientific projects benefit from keeping parameters — hyperparameters, paths, thresholds — out of the scripts that consume them, as a single source of truth.

CONFIG = load_configs() collects every YAML under <project-root>/config/ into one nested DotDict. Parameters are then accessible as CONFIG.YAML_FILE_NAME.FIELD_NAME.

UPPER_CASE normalisation. YAML filenames and field names are recognised in UPPER_CASE, following Python's convention for user-defined parameters. model.yaml with hidden_dim: 256 lands at CONFIG.MODEL.HIDDEN_DIM regardless of source casing.

Conflict handling. When an UPPER/lower pair collide (e.g. MODEL.yaml next to model.yaml, or HIDDEN_DIM next to hidden_dim), the UPPER variant is prioritised and a UserWarning is emitted pointing at the conflict.

/path/to/project/
├── config/
│   ├── PATHS.yaml                       # DATA_DIR: /data/experiment_01
│   ├── PREPROCESS.yaml                  # SAMPLE_RATE: 1000, BANDPASS: [0.5, 40]
│   ├── MODEL.yaml                       # HIDDEN_DIM: 256, DROPOUT: 0.3
│   └── IS_DEBUG.yaml                    # IS_DEBUG: true
└── scripts/
    └── xxx/
        └── filename.py                  # CONFIG = sio.load_configs()
                                         #   CONFIG.MODEL.DROPOUT       → 0.3
                                         #   CONFIG.PREPROCESS.SAMPLE_RATE → 1000

CONFIG = sio.load_configs()           # loads ./config/*.yaml
CONFIG.PREPROCESS.SAMPLE_RATE            # 1000

# Debug mode: DEBUG_ prefixed keys override their counterparts
# In MODEL.yaml: { HIDDEN_DIM: 256, DEBUG_HIDDEN_DIM: 32 }
CONFIG = sio.load_configs(IS_DEBUG=True)
CONFIG.MODEL.HIDDEN_DIM                  # 32 (debug value promoted)

Debug mode for parameters

When debugging or developing, flipping parameters speeds up iteration. Any DEBUG_* sibling overrides its non-debug counterpart at load time (e.g. CONFIG.MY.DEBUG_PARAM replaces CONFIG.MY.PARAM), so a single IS_DEBUG.yaml flips the whole project between production and debug values.

Equivalent triggers — these three all enable debug mode: IS_DEBUG.yaml with IS_DEBUG: true, load_configs(IS_DEBUG=True), or running under CI=True.

4. Linter for Migration and Hooks

scitex-io ships 14 IO-specific (STX-IO001..014) and 5 path-handling (STX-PA001..005) lint rules. They are detected automatically by scitex-dev's linter, which is already a hard dependency of scitex-io — no extra install needed.

scitex-dev linter check-files src/           # lint a tree
scitex-dev linter list-rules --category io   # show live rule definitions

Rule reference (STX-IO001..014 + STX-PA001..005)

Rule	Severity	Trigger
STX-IO001	warning	np.save / savez / savez_compressed / savetxt → use sio.save()
STX-IO002	warning	np.load / loadtxt / genfromtxt → use sio.load()
STX-IO003	warning	pd.read_csv / parquet / excel / hdf / pickle / json / feather / orc / table → use sio.load()
STX-IO004	warning	df.to_csv / parquet / excel / hdf / pickle / json / feather / html / orc → use sio.save()
STX-IO005	warning	pickle.dump / dumps / load / loads (incl. cPickle) → use sio.save()/load()
STX-IO006	warning	json.dump / dumps / load / loads → use sio.save()/load()
STX-IO007	warning	.savefig(...) → use sio.save(fig, path) for metadata embedding
STX-IO008	warning	torch.save / load → use sio.save()/load()
STX-IO009	warning	joblib.dump / load → use sio.save()/load()
STX-IO010	warning	yaml.dump / safe_dump / dump_all / load / safe_load / full_load → use sio.save()/load()
STX-IO011	warning	scipy.io.savemat / loadmat → use sio.save()/load()
STX-IO012	warning	cv2.imread / imwrite, PIL.Image.open, plt.imsave / imread, imageio.* → use sio.save()/load()
STX-IO013	warning	h5py.File(...) → use sio.save()/load() for HDF5
STX-IO014	warning	sio.save / load called with an extension that has no registered handler — register one with register_saver/register_loader
STX-PA001	warning	Absolute path passed to sio.* — prefer relative for reproducibility
STX-PA002	warning	open(...) → use sio.save()/load() for auto-logging
STX-PA003	info	os.makedirs / mkdir — sio.save() auto-creates directories
STX-PA004	warning	os.chdir(...) — scripts should run from project root
STX-PA005	info	Relative path missing ./ prefix — use ./file.ext for explicit intent

Claude Code Integration as a Hook

Wire scitex-io's lint rules into Claude Code so every Edit / Write to a Python file is checked automatically — errors block the turn, warnings surface as feedback.

Reference implementation: examples/scitex_io_lint.sh — a self-contained PostToolUse hook (~15 lines) that you can copy into ~/.claude/hooks/post-tool-use/ (or into your project's .claude/hooks/).

1. Install the hook script:

cp examples/scitex_io_lint.sh ~/.claude/hooks/post-tool-use/
chmod +x ~/.claude/hooks/post-tool-use/scitex_io_lint.sh

2. Wire it up — add to ~/.claude/settings.json (or <project>/.claude/settings.json for project-scoped):

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write|MultiEdit",
        "hooks": [
          { "type": "command",
            "command": "~/.claude/hooks/post-tool-use/scitex_io_lint.sh" }
        ]
      }
    ]
  }
}

After that, every time Claude Code edits a .py file, an STX-IO001..014 / STX-PA001..005 error blocks the turn and Claude sees the rule message inline — agents converge on the canonical sio.save() / sio.load() patterns instead of np.save / pd.read_csv / pickle.dump / ….

5. Etc.

Glob, parse, cache

paths = sio.glob("data/**/*.csv")                        # natural sort: 1, 2, 10
paths = sio.glob("results/{exp1,exp2}/*.npy")            # brace expansion
paths, parsed = sio.parse_glob("sub_{id}/ses_{session}/*.vhdr")
# parsed = [{'id': '001', 'session': 'pre'}, ...]

dfs = sio.load("results/*.csv")                          # list of DataFrames
data = sio.load("large.hdf5"); data = sio.load("large.hdf5")  # 2nd call: cache hit

Embed provenance into figures (embed_metadata)

sio.embed_metadata("figure.png", {
    "experiment": "exp_042", "model": "resnet50",
    "accuracy": 0.94, "timestamp": "2026-03-11",
})
meta = sio.read_metadata("figure.png")
meta["experiment"]              # "exp_042"

Supports PNG (tEXt), JPEG (EXIF), SVG (XML metadata), PDF (XMP).

Four Interfaces

Python API ⭐⭐⭐

from scitex_io import save, load, list_formats, register_saver, register_loader
from scitex_io import load_configs, DotDict
from scitex_io import embed_metadata, read_metadata, has_metadata

save(obj, "path.ext")        # Save any object
data = load("path.ext")      # Load any file
fmts = list_formats()        # Show all registered formats
cfg  = load_configs()        # Load ./config/*.yaml as DotDict
embed_metadata("fig.png", d) # Embed provenance into figure

Full API reference

CLI Commands ⭐

scitex-io --help-recursive          # Show all commands
scitex-io info                      # Show registered formats
scitex-io configs                   # Load and display project configs
scitex-io configs -d ./my_configs   # Custom config directory
scitex-io configs --json            # Output as JSON
scitex-io list-python-apis -vv      # List Python APIs with signatures
scitex-io --version                 # Show version
scitex-io mcp start                 # Start MCP server
scitex-io mcp doctor                # Check MCP health
scitex-io mcp list-tools -vv        # List MCP tools with parameters

Full CLI reference

MCP Server ⭐

AI agents can save, load, and discover formats autonomously.

Tool	Description
io_list_formats	List all registered save/load formats
io_load / io_save	Load / save data in any supported format
io_load_configs	Load YAML project configurations
io_register_info	Show how to register custom formats
io_glob / io_parse_glob	Natsorted globbing with {placeholder} parsing
io_get_loader / io_get_saver	Look up the registered handler for an extension
io_read_metadata / io_has_metadata / io_embed_metadata	Image provenance metadata
io_get_cache_info / io_clear_load_cache / io_configure_cache	Load-cache management
io_explore_h5 / io_explore_zarr	Print group/dataset trees
io_has_h5_key / io_has_zarr_key	Cheap existence checks
io_json2md	Render JSON as Markdown
io_skills_list / io_skills_get	Discover and fetch skill pages

scitex-io mcp start

Full MCP specification

Skills ⭐⭐

Skills provide structured documentation that AI agents can query to discover package capabilities, API signatures, and usage patterns.

scitex-io skills list              # List available skill pages
scitex-io skills get save-and-load # Get detailed save/load documentation
scitex-io skills get glob          # Get glob/parse_glob patterns
scitex-io skills get supported-formats  # Get all format tables

Skill	Content
save-and-load	Core API, path routing, symlinks, use_caller_path
centralized-config	load_configs(), DotDict, DEBUG_ override
metadata-embedding	Provenance in PNG/JPEG/SVG/PDF
cache	Load caching, reload, flush
glob	Pattern matching with natural sort and parsing
linting-rules	STX-IO001–007 lint rules
supported-formats	All 30+ format tables
path-resolution	Auto save-path routing, scitex.path utilities

Also available via MCP: io_skills_list() / io_skills_get(name).

Part of SciTeX

scitex-io is part of SciTeX. Install via the umbrella with pip install scitex[io] to use as scitex.io (Python) or scitex io ... (CLI).

import scitex

@scitex.session
def main(CONFIG=scitex.INJECTED):
    data = scitex.io.load("input.csv")     # auto-tracked by clew
    result = process(data)
    scitex.io.save(result, "output.csv")   # auto-tracked by clew
    return 0

scitex.io delegates to scitex_io — they share the same API and registry.

The SciTeX system follows the Four Freedoms for Research below, inspired by the Free Software Definition:

Four Freedoms for Research

0. The freedom to run your research anywhere — your machine, your terms.
1. The freedom to study how every step works — from raw data to final manuscript.
2. The freedom to redistribute your workflows, not just your papers.
3. The freedom to modify any module and share improvements with the community.

AGPL-3.0 — because we believe research infrastructure deserves the same freedoms as the software it runs on.

---