decoui

Decorator-driven GUI framework for Python. Annotate your methods — decoui generates a full PySide6 desktop app automatically.

---

Features

• Zero UI code — decorate a class and its methods, call gui_main()
• Native type mapping — str, int, float, bool, list, dict, Enum → widgets automatically
• Form assist — autocomplete, cascading fill, and lazily evaluated defaults
• Async execution — every tool runs in a thread; stdout and logging are captured in real time
• Execution history — every run is persisted to SQLite with parameters, logs, and status
• Parallel tool tabs — keep multiple tool pages open and switch between running tasks
• Persistent layout — sidebar width is restored from the application database
• Replay — restore any past run's parameters with one click
• Settings store — store("ns")[key] = value, namespaced and persisted in the same database
• Help panel — a reference built from the docstrings the tools already carry, with cross-references, tabs and back/forward
• Themes — four built in, any number as JSON files; changed from Settings without a restart
• Translatable interface — decoui's own text ships as one JSON catalogue per language

---

Installation

pip install decoui

Requires Python 3.10+.

decoui depends on PySide6-Essentials (6.6+), not the full PySide6 metapackage. It only ever imports QtCore, QtGui, QtWidgets and QtTest, all of which live in Essentials — so the extra PySide6-Addons payload (Charts, Multimedia, WebEngine, Qt3D, …) is skipped. Measured on Windows / PySide6 6.11: 637 MB → 202 MB.

If your own tool methods need an Addons module, depend on it explicitly:

pip install decoui PySide6-Addons   # or simply: pip install decoui PySide6

Both coexist with decoui — PySide6 is just PySide6-Essentials + PySide6-Addons.

Upgrading an existing environment — all three distributions unpack into the same PySide6/ directory, so uninstalling PySide6 / PySide6-Addons also deletes shared libraries that Essentials still needs, leaving ImportError: DLL load failed while importing QtWidgets. Force a clean reinstall rather than relying on the uninstall:

pip install --force-reinstall PySide6-Essentials
# uv: uv sync --reinstall-package pyside6-essentials

Fresh installs are unaffected.

---

Quick Start

import logging
from decoui import tool, toolset, gui_main

@toolset(label="Text Tools", tags=["text"])
class TextTools:

    @tool(label="Count Characters",
          description="Count chars, words and lines.",
          placeholders={"content": "Paste text here…"})
    def count(self, content: str = "") -> str:
        words = len(content.split())
        logging.info("words=%d", words)
        return f"{len(content)} chars / {words} words"

if __name__ == "__main__":
    gui_main(title="My Tools")

gui_main() auto-discovers every @toolset class in the caller's global scope — no explicit registration needed.

---

API Reference

@toolset

Applied to a class. Groups its @tool methods under one sidebar entry.

@toolset(
    label="CSV Tools",           # required — sidebar display name
    tags=["file", "batch"],      # optional — used by the tag filter bar
    description="...",           # optional — tooltip on hover in the sidebar
)
class CsvTools:
    ...

Parameter	Type	Description
label	str	Required. Display name in the sidebar.
tags	list[str]	Tag filter bar labels. Multiple tags = AND filter.
description	str	Shown as a tooltip when hovering the toolset in the nav tree.

@tool

Applied to a method inside a @toolset class.

@tool(
    label="Merge Files",
    description="Merge multiple files into one.",
    placeholders={"files": "one path per line"},
    confirm=True,     # show confirmation dialog before running
    timeout=120,      # cancel after N seconds (None = unlimited)
)
def merge(self, files: list, output: str = "out.txt") -> str:
    ...

Parameter	Type	Default	Description
label	str	required	Tool display name.
description	str	""	Shown in a bordered box below the title.
placeholders	dict[str,str]	{}	Placeholder text per parameter name.
labels	dict[str,str]	{}	Form label per parameter name. Defaults to the parameter name.
confirm	bool	False	Show a Yes/No confirmation dialog before running.
timeout	int|None	None	Execution timeout in seconds.
completions	dict[str,list|callable|str]	{}	Autocomplete candidates per parameter. See Form Assist.
cascade	dict[str,callable|str]	{}	Fill other parameters when this one changes.
defaults	dict|callable|str	None	Initial form values, evaluated when the page opens.
completion_debounce_ms	int	250	Idle time before a dynamic completions callback runs.
on_cancel	callable|str	None	Cleanup to run when the tool is cancelled. See Cancellation.

Every key in placeholders, labels, completions, cascade and defaults must name a real parameter of the method, and placeholders/labels values must be strings. A mistake raises at startup, when the toolset tree is built, naming the tool and listing the parameters it does have.

gui_main

gui_main(title="My App", db_path="~/.myapp/history.db", on_startup=connect_backend)

Parameter	Type	Default	Description
title	str	"decoui"	Window title.
db_path	str|Path|None	~/.decoui/history.db	SQLite database path for execution history and application settings.
on_startup	callable|None	None	Application-wide setup, run once before any toolset is created and before the window appears. See Startup.
toolsets	Sequence[type]|None	None	The @toolset classes to load. When omitted, every one visible in the calling namespace is discovered.
theme	str|None	None	Default theme id. A theme the user picked in Settings wins over it. See Themes.
theme_dir	str|Path|None	~/.decoui/themes	Directory scanned for user-supplied theme files.
language	str|None	None	Code for the language decoui's own interface is drawn in — Run, Stop, the history columns. A language the user picked in Settings wins over it. A tool's own label, description and docstring are never translated. Defaults to English.

By default gui_main() scans the caller's namespace, so a toolset has to be imported and look used:

from mytools.restore import RestoreTools  # noqa: F401 -- discovered by gui_main()

gui_main(title="My App")

Naming them explicitly drops the pragma and the comment that has to explain it:

from mytools.restore import RestoreTools

gui_main(title="My App", toolsets=[RestoreTools])

The list controls what loads, not the order it appears in: the sidebar is always sorted by label.

Startup

Full guide: docs/startup-lifecycle.md — read it before loading anything at startup.

Startup runs in a fixed order, so anything loaded early is available later:

1. db_path is applied and the database is initialised
2. the toolset tree is built and validated
3. gui_main(on_startup=...) runs — application-wide work
4. every @toolset class is instantiated
5. each instance's own on_startup() method runs, if it defines one
6. the window is shown and the event loop starts

The recommended pattern is to declare state in __init__ and load it in on_startup():

@toolset(label="Deploy Tools")
class DeployTools:

    def __init__(self):
        self.config = {"env": "staging"}          # declare, with fallbacks

    def on_startup(self):
        self.config = json.loads(                  # load, before the UI is live
            CONFIG_PATH.read_text(encoding="utf-8")
        )

    @tool(label="Deploy", defaults="load_defaults")
    def deploy(self, env: str = "") -> None:
        ...

    def load_defaults(self) -> dict:
        return {"env": self.config["env"]}         # reads self at call time

Splitting the two matters when loading fails: __init__ still succeeded, so the toolset opens with its declared fallbacks instead of disappearing.

Use the gui_main(on_startup=...) argument for work that belongs to no single toolset — a shared connection, an auth token, a warm cache. It runs before any instance exists, so it cannot write to self; give the toolset its own on_startup() method for that. The application hook always runs before every toolset hook.

Two rules for both hooks

• The event loop is not running yet. QTimer does not fire, queued signals are not delivered, and QThreadPool results never arrive — waiting on a background thread here deadlocks startup. To report a problem, just raise; decoui shows it in a dialog once the window is up.
• They block the window. Both run synchronously before show(), so however long they take is time the user spends looking at nothing. Anything slow belongs in the tool body, where it gets a thread, a progress bar, and a Stop button.

Can the decorator see self?

Decorator arguments are evaluated at import time, when no instance exists — so this is impossible:

@tool(label="Deploy", defaults={"env": self.config["env"]})   # NameError: self

Pass a method name instead. decoui resolves it to a bound method when the tool page is built, and the method reads self at call time — after __init__ and on_startup() have run:

@tool(label="Deploy",
      defaults="load_defaults",          # self.config is available inside
      completions={"service": "search_services"},
      cascade={"service": "describe_service"})

This works for defaults, completions, and cascade alike.

When startup fails

A failure in step 3, 4, or 5 does not stop the application. Every failure is collected and reported in one dialog over the main window, listing what failed with the full traceback behind the Details button, and everything that did load stays usable.

The one exception is step 4: a class whose __init__ raises produces no object, so that toolset is skipped and its tools do not appear. That is the practical reason to keep __init__ trivial and do the real loading in on_startup().

---

Type → Widget Mapping

Python annotation	Widget	Notes
str	QLineEdit	Single-line text.
int	QSpinBox	Integer, full int range.
float	QDoubleSpinBox	4 decimal places.
bool	QCheckBox	Checked / unchecked.
list	QTextEdit	One item per line or comma-separated.
dict	QTextEdit	JSON input; parsed with json.loads then ast.literal_eval. Raises on invalid input.
Enum subclass	QComboBox	Dropdown of enum members.
pathlib.Path	QLineEdit + buttons	Text field with File... (file picker) and Folder... (directory picker) buttons. The selected path is passed as a pathlib.Path to the method.

• Required parameters (no default) are marked with a red * in the form label.
• Optional[X] is unwrapped to X.
• Annotated[X, ...] maps on X; the metadata does not affect widget choice.
• Default values are pre-filled into widgets automatically.

Field metadata with Annotated

F attaches a label and placeholder to the type itself, so a field shared across tools carries its wording with it instead of repeating it in every @tool:

from pathlib import Path
from typing import Annotated
from decoui import F, tool, toolset

DumpFile = Annotated[Path, F(label="Dump file", placeholder="Pick a .dump / .tar")]

@toolset(label="Backup")
class Backup:

    @tool(label="Restore")
    def restore(self, archive: DumpFile) -> None: ...

    # One tool can reword the shared field; the decorator wins.
    @tool(label="Verify", labels={"archive": "Archive to check"})
    def verify(self, archive: DumpFile) -> None: ...

Resolution order for both label and placeholder:

@tool(labels=…) / @tool(placeholders=…) → Annotated[…, F(…)] → parameter name (label) or empty (placeholder)

Adding F to an annotation is purely additive: widget selection, required-field marking and value conversion all read the bare type underneath, so existing code keeps working unchanged.

Note — with from __future__ import annotations, annotations are strings at runtime and decoui resolves them with typing.get_type_hints(). An alias like DumpFile must therefore be importable at runtime; putting the import under if TYPE_CHECKING: makes resolution fail, and decoui falls back to the raw strings, losing the F metadata along with the type mapping.

pathlib.Path example

import pathlib
from decoui import tool, toolset

@toolset(label="File Tools")
class FileTools:

    @tool(label="File Info", placeholders={"path": "Select a file or folder…"})
    def file_info(self, path: pathlib.Path) -> None:
        import logging
        p = pathlib.Path(path)
        logging.info("size: %d bytes", p.stat().st_size)
        logging.info("resolved: %s", p.resolve())

---

Form Assist

Three opt-in @tool arguments make the parameter form react to what the user types.

_CATALOG = {
    "web-frontend": {"version": "2.4.1", "owner": "frontend-team"},
    "auth-service": {"version": "3.0.2", "owner": "identity-team"},
}

@toolset(label="Deploy Tools")
class DeployTools:

    @tool(
        label="Deploy Service",
        completions={
            "env": ["production", "staging", "development"],   # static list
            "service": "search_services",                       # method on this class
        },
        cascade={"service": "describe_service"},                # fill other fields
        defaults={"env": "staging"},                            # initial values
    )
    def deploy(self, service: str = "", env: str = "",
               version: str = "", owner: str = "") -> None:
        ...

    def search_services(self, text: str) -> list[str]:
        return [name for name in _CATALOG if text.casefold() in name]

    def describe_service(self, value: str, form: dict) -> dict:
        return dict(_CATALOG.get(value, {"version": "", "owner": ""}))

completions — autocomplete

Value	Behaviour
list[str]	Fixed candidates, filtered by Qt (case-insensitive, substring match).
callable	Re-evaluated as the user types, debounced. The returned list replaces the popup contents; your callback controls matching and ordering.
str	Name of a method on the toolset class, so the callback can use self.

Clicking into the field opens the popup right away, before anything is typed — a static list shows every option, and a callback is called with an empty text so it can return its default set. The list is not narrowed by a value already in the field, since re-entering one is usually a prelude to changing it.

Supported on str and pathlib.Path parameters only. Declaring it elsewhere raises at startup.

cascade — fill other fields

Keyed by the source parameter. The callback returns {target_name: value}, or None to change nothing. It fires when the source is committed (focus loss, Enter, picking a candidate, toggling a checkbox), not on every keystroke, and is skipped when the value did not actually change.

Targets are always overwritten, including fields the user typed into — the callback receives form with every current value, so it can return one unchanged to preserve it. A written target may itself be a cascade source; chains are capped at 5 levels and cycles terminate after one round. Replay never triggers cascades, so restored parameters survive.

defaults — lazy initial values

Signature defaults (def deploy(self, env: str = "staging")) are evaluated at import time. Use defaults= when the initial value must be read after the application has started — for example from ~/.decoui/history.db, which does not exist until gui_main() runs.

@tool(label="Deploy", defaults=lambda: {"env": get_setting("deploy.env", "staging")})

Accepts a dict, a callable taking nothing or the current form, or a method name. Runs once per tool page.

Callback rules

• Completion and cascade callbacks run on a background thread. They must not touch Qt objects. Keep them to a lookup: query, return data.
• Callbacks may take 0, 1, or 2 positional arguments — (text, form) for completions, (value, form) for cascade.
• Any exception is caught, reported as a WARNING in the output console, and never aborts the form.
• Results from superseded requests are discarded, so a slow lookup cannot overwrite a newer one.

---

Output & Logging

Tool methods can use print() and the standard logging module. Both are captured and rendered in the output console with colour coding:

Source	Colour
print / stdout	White
logging.DEBUG	Gray
logging.INFO	Phosphor Green
logging.WARNING	Yellow
logging.ERROR	Red
logging.CRITICAL	Bold Red

Return values from tool methods are not displayed in the GUI. Use logging or print for any output you want users to see.

run_process — calling an external program

from decoui import run_process

result = run_process(["pg_dump", "-d", "app"], check=True)

Runs the program, streams its output into the console line by line, and lets Stop kill it — and everything it spawned — without the tool declaring anything. Plain subprocess does neither. See Cancellation.

Progress

Long, quiet work looks indistinguishable from a hang. progress() drives the page's progress bar and the status text next to it:

from decoui import progress, tool, toolset

@tool(label="Dump")
def dump(self, schema: str) -> None:
    for index, table in enumerate(tables):
        progress(index, len(tables), f"dumping {table}")
        dump_table(table)
    progress(len(tables), len(tables), "done")

Argument	Meaning
done	Units completed so far.
total	Total units, or 0 when unknown — the bar stays indeterminate and only the message updates.
message	Short status text. Replaces "Running…" while set.

Calls closer than 100 ms apart are dropped so a tight loop cannot flood the GUI; the final call (done >= total) is always delivered.

Outside a running tool — when you instantiate the toolset and call the method directly, as in a test — progress() does nothing at all. It prints nothing and raises nothing, so a tool stays callable as a plain method.

---

Tool Page Buttons

Button	Action
Run	Type-coerce parameters, run the method in a background thread.
Reset	Expand the parameter panel and clear the output console. Parameters are kept.
Stop	Request cancellation of the running task.
Replay	Open the History panel pre-filtered to this tool's past runs.
Copy	Copy current console output to clipboard.
View Log	Open the current console output in a resizable log viewer window.

---

Cancellation

Stop is the hardest promise decoui makes, because Python cannot keep it on its own.

Cancellation works by injecting an exception into the worker thread, and CPython raises it at the next bytecode boundary. A thread parked in a C call never reaches one. proc.wait(), socket.recv(), a long time.sleep() — the exception stays pending until that call returns by itself, and the tool's finally does not run either.

So whatever is holding the thread has to be released from outside. Everything below follows from that one fact.

If your tool shells out, use run_process

from decoui import run_process, tool, toolset

@toolset(label="Database")
class RestoreTools:

    @tool(label="Restore")
    def restore(self, archive: Path) -> str:
        result = run_process(["pg_restore", str(archive)])
        if result.cancelled:
            return "Stopped."
        return f"pg_restore exited {result.returncode}."

No handle to keep, no on_cancel to declare. run_process registers the child with the running tool, so Stop kills it and everything it spawned.

It also streams the child's output into the page console line by line, which plain subprocess cannot: decoui replaces sys.stdout with an object that has no fileno(), so a child left to inherit stdout writes past the console rather than into it.

result.returncode	the child's exit status
result.output	everything it wrote, already printed
result.cancelled	True only when decoui killed it
check=True	raise ProcessError if the program fails — a cancelled run never raises, because that status is decoui's doing
shell=True	refused: a shell is what leaves an unkillable grandchild behind. Pass a list

subprocess.run() cannot be made stoppable, with or without a hook — it keeps its handle to itself, so there is nothing for a hook to terminate. That is the shape a real incident had: Stop appeared to work and the dump kept running.

Measured, not assumed: every claim here is a test in tests/test_stop_external_process.py, driving the real engine against a real child process. Full matrix in docs/cancelling-a-run.md.

If your tool blocks on something else, declare on_cancel

For a socket read, a database driver, a lock — anything run_process cannot reach — the hook is still the only way. It runs on the GUI thread, before the interrupt is injected, while the worker is still blocked:

@tool(label="Query", on_cancel="stop")
def query(self) -> None:
    self.conn = driver.connect(...)
    self.conn.execute(long_query)

def stop(self) -> None:
    conn = getattr(self, "conn", None)
    if conn is not None:
        conn.cancel()

Three properties the hook must have:

• Idempotent. A Stop that arrives before the tool has assigned the state the hook reads does nothing, and a second Stop is the user's only way out of that — hence getattr(self, "conn", None) rather than self.conn.
• Fast. It blocks the GUI. Slow cleanup belongs in the tool's own finally.
• Concurrent with your tool. It is on the GUI thread while the body is still on the worker thread, so do not touch state the body is writing.

A hook that raises is reported as an ERROR line in that run's log; cancellation still completes.

What else to know

except Exception does not catch it. The injected _WorkerCancelled derives from BaseException, so an ordinary handler lets it pass straight through. That is deliberate — cleanup written as except Exception must not be able to swallow a cancellation and carry on.

A cancelled run is recorded as cancelled, never success, and the return value is dropped. A tool that returns normally after Stop was pressed is still recorded as cancelled: the user asked for it to stop, and a partial result must not look like a whole one.

Do not trust proc.returncode after a cancellation if you manage a child yourself. The injection can land between waitpid() returning and Popen recording the status, so it may stay None even though the child is gone. Ask the OS.

timeout= uses this same path — from the tool's side a timeout and a Stop press are the same event.

---

Themes

A theme sets colours, corner radii, border widths and fonts -- per part, not per application: the sidebar, the tool list inside it, input fields, buttons, tabs, tables and the output console each have their own tokens, so a theme can put a dark panel behind the tool list while the rest of the window stays light. It cannot add or move widgets.

id	Name	Look
light	Light	The default. Rounded, blue accent.
cockpit	Cockpit Panel	Military instrument panel: olive-grey chassis, green accent, hard edges.
nasa	Mission Control	Warm off-white console: square corners, deep green accent, monospace.
jp-industrial	Industrial 1980s	Japanese workstation: beige chassis, burnt-orange accent, square.

The three panel themes are recolours and re-geometries of the same interface, not replicas of a hardware console: decoui has no header band, status lamps or bezel screws to dress, and a theme cannot add any.

gui_main(title="Ops", theme="light", theme_dir="~/.decoui/themes")

Parameter	Meaning
theme	The application's default theme id. A theme the user picks in Settings wins over it.
theme_dir	Where user themes are read from. Defaults to ~/.decoui/themes. A missing directory is fine and is never created.

Writing one

Drop a .json file into the theme directory. Nothing needs rebuilding, and the new theme appears the next time the application starts.

{
  "version": 1,
  "id": "brand",
  "name": "Brand",
  "extends": "light",
  "colors": {
    "accent": "#0f766e",
    "accent.soft": "#ccfbf1",
    "bg.tree": "#0d1a0b",
    "text.on_sidebar": "#cfd8c3"
  },
  "shape": { "shape.radius_control": 2 },
  "font": { "mono_family": ["JetBrains Mono", "monospace"], "small_size_pt": 8 }
}

The token groups:

Group	Count	Covers
bg.*	21	every surface separately -- app, page, top bar, tabs, sidebar, tool list, fields, buttons, table, console
text.*	16	one ink per place text sits, including text.on_sidebar (a dark tool list), text.on_topbar (a dark top bar), and text.on_running / on_success / on_danger / on_neutral so a bright Run button can take dark text while Stop stays dark and takes light text
border.*	10	panels, fields, buttons, tabs, the cap on the current tab, focus, and the console's frame
console.*	12	a formatted line is inked in three parts -- console.timestamp, then console.tag.<level> for the level itself and console.body.<level> for the message. console.plain covers a line that carries no level at all, such as raw print() output. Set a level's tag and body to one value to tint the whole line
accent · running · success · danger · neutral · scrollbar.*	11	selections and the semantic fills. running is separate from accent because accent also fills every checked button: a theme that wants its toggles and its Run button in one colour would otherwise get a Running badge identical to the Done one
shape.*	14	six corner radii, and border widths and styles grouped the way the colours are -- _panel / _control / _field, plus _emphasis and _focus -- so a theme can bevel its buttons without bevelling its tables. Qt's outset / inset / ridge / groove draw a bevel from the border colour, which is as close to a raised panel as a flat format gets; double needs a width of at least 3 before two lines fit. Set the radii to 0 to square everything off.
font.*	8	family / size_pt / letter_spacing, mono_family / mono_size_pt for the console, title_size_px / small_size_pt for headings and secondary controls, and uppercase to render tags, tabs and buttons in capitals

Font families are stacks: Qt falls through them in order, so end every one with a generic family (monospace, sans-serif) or the theme lands on Qt's default wherever its preferred face is missing.

• id is the stable key — it is what gets saved when the user selects the theme, so renaming name never loses their choice.
• extends starts from a built-in theme and overrides only the keys you list. Without it, every token must be present.
• extends always resolves against the theme decoui ships, even if another file has taken over that id — so one theme can never quietly re-base another.
• Giving your theme the id of a built-in replaces it, with no warning — that is how you re-skin light or nasa. Two of your own files claiming one id is reported instead, because one of them silently loses.
• Colours are #rrggbb only. There are no gradients: every token is one flat colour, so metallic and bevelled looks are out of reach.
• Setting the shape.radius_* tokens to 0 squares the whole interface off.
• font.uppercase changes only how labels are drawn -- the underlying strings are untouched, so a tab's title still matches its tool id.

Run python -c "import decoui.theme as t; print(sorted(t.COLOR_TOKENS))" for the full token list.

When a theme is broken

A theme is presentation, so a bad one never stops the application:

Situation	What happens
One file is invalid	It is skipped; every other theme still loads
The selected theme is invalid or missing	The light theme is used instead
Either of the above	The application starts, and reports it in the startup dialog

A selected theme that has gone missing is not un-selected — the file may be absent only on this machine, and the choice takes effect again once it returns.

Changing theme

The gear button at the top right opens Settings, which lists every theme available -- built-in and user-supplied alike -- and starts on the one currently in effect.

A new theme is applied as soon as the dialog closes, to every window that is open. Nothing is rebuilt: a tool that is running goes on running, the forms keep what was typed into them, and output already printed is re-inked in the new colours. The one thing that is lost is the console's scroll position, which returns to the newest line.

The interface language, chosen in the same dialog, is the exception: it takes effect on the next launch. Text is read as each widget is built, in far more places than colour is, and there is no equivalent of the application stylesheet to catch the rest. The dialog says which is which before you choose.

---

Help Panel

The ? button in the top bar opens a reference for everything loaded in the session. Nothing has to be declared for it: a tool's page is built from the docstring the method already carries — its summary, the prose under it, and the Args:, Returns: and Raises: sections of Google style.

Write the docstring for the person using the tool, not for the person reading the file. Notes about which annotation produces which widget belong in # comments above the method; comments are not collected, so the two audiences stay separated.

Alongside the tools, the panel carries decoui's own guide — how history works, what Replay actually replays, how to change theme. That part ships with decoui and is translated with the rest of the interface.

Each page opens in its own tab, the way the main window opens a tool, and the arrows above the tabs walk back and forward through the pages visited.

Writing help

Docstrings and guide pages are Markdown: headings, fenced code blocks, tables, ordered and nested lists, blockquotes. Two departures, because the source is sometimes a Python docstring:

• ``literals`` in double backticks are accepted alongside single, since that is what reST — and therefore a Python docstring — uses;
• indented code blocks are not recognised. A docstring's indentation is an artefact of where it sits in the file. Fence code instead.

Cross-references have their own syntax, [[key]], deliberately not Markdown's link syntax:

[[MyTools.encode]]                one tool
[[MyTools]]                       a whole toolset
[[guide.themes]]                  one of decoui's own pages
[[guide.themes|the theme page]]   with your own text

The target is a key, never a title or a file name:

Key	Page
guide	decoui's own contents page
guide.<slug>	one of decoui's guide pages
ClassName	a toolset
ClassName.method	a tool

Keys do not change when a page is translated, so one written link works in every language. A reference that does not resolve — a tool the application did not load, a typo — renders as its own text with no link on it. Which tools exist is up to the application, so help cannot be written against a fixed set, and a dead link is worse than the sentence without it.

Keeping references out of [text](target) is what lets that form mean what it means everywhere else: [text](https://…) is an ordinary link and opens in the reader's browser. Only http, https and mailto are followed.

Link colour comes from the theme's text.link token.

Help in a file

When a tool's help outgrows its docstring, or wants translating, point at a Markdown file:

@tool(label="Deploy", help="doc/deploy.md")
def deploy(self, service: str = "web") -> str:
    ...

The path is relative to the module the toolset class is defined in, and decoui inserts a language directory into it — doc/<language>/deploy.md, falling back to doc/en/deploy.md and then doc/deploy.md.

The file replaces the prose and nothing else: the summary, the parameter table and Returns still come from the docstring, because they describe the signature and a file beside the module cannot be checked against it.

Full reference: docs/help-authoring.md.

Translating your own tools

decoui translates its own interface from catalogues it ships. The labels, descriptions and field text you write into @toolset and @tool are strings it has never seen, so they get a catalogue of your own:

gui_main(toolsets=[...], i18n_dir="i18n")     # i18n/ja-JP.json, i18n/zh-CN.json

Keys are ClassName and ClassName.method — the same keys a cross-reference uses. Every field is optional; an absent key leaves the string your source wrote, so a half-finished catalogue gives a half-translated interface rather than a broken one.

{
  "OpsTools.restore": {
    "label": "復元",
    "brief": "アーカイブから復元します。",
    "returns": "復元された内容。",
    "params": { "archive": { "label": "アーカイブ", "brief": "読み込むアーカイブ。" } }
  }
}

This cannot be done in the decorator: its arguments are evaluated at import, before gui_main() settles the language. decoui substitutes later, in build_tree(), which is what reaches the sidebar, the tabs, the forms, the Help panel and the history all at once.

decoui can write the starting point for you — a catalogue holding every translatable string your code declares, and a Markdown file per tool holding its prose. Both live behind a hidden switch, off in anything you ship.

Full reference, including how to turn that switch on: docs/translating-an-application.md.

---

Remembering Things Between Runs

A tool often has one thing worth keeping — the environment last deployed to, the folder last exported into, whether the verbose flag was on. store() gives you a namespaced key/value store, persisted in the same SQLite database as the run history.

from decoui import store, tool, toolset

@toolset(label="Deploy")
class DeployTools:

    def load_defaults(self) -> dict:
        return {"env": store("deploy").get("last_env", "staging")}

    @tool(label="Deploy Service", defaults="load_defaults")
    def deploy(self, service: str, env: str) -> None:
        """Deploy a service.

        Args:
            service: What to deploy.
            env: Where to deploy it.
        """
        store("deploy")["last_env"] = env      # inserted if new, replaced if not

It behaves as a mapping:

s[key] = value / s.set(key, value)	write; there is nothing to register first
s[key]	read, KeyError when absent
s.get(key, default)	read with a fallback
del s[key] / s.delete(key)	remove; removing what was never there is fine
key in s, len(s), list(s)	the usual
s.keys(), s.items()	everything in this namespace, sorted

The namespace

store("deploy") and store("deploy") anywhere else in the application reach the same rows — that is how a setting shared by every tool is shared. Omit the name and you get app, for an application that only needs one.

ui and decoui are refused: they hold the chosen theme, the interface language and the sidebar width. An application writing there would be changing the user's settings rather than its own.

A namespace may not contain a dot, because the dot is what separates it from the key. Keys may contain anything.

Values

Anything JSON can carry — str, int, float, bool, None, and lists and dicts of those — and it comes back as the type it went in as. A value JSON cannot carry raises TypeError where you wrote it, rather than being coerced to a string that fails somewhere else later.

None is a stored value, not an absence: key in store still reports True.

Tuples come back as lists. JSON has no tuple.

What it is not

A settings store, not an application database: small values, one at a time, no queries and no relations. A tool with real data of its own should open its own file.

Reads and writes go straight to the database — there is no cache — and every call opens and closes its own connection, so this is safe to use from a tool body, which runs on a worker thread.

Clearing the run history does not clear these: clear_all_records() leaves the settings table alone.

---

Execution History

Every run is stored in SQLite. The History panel (sidebar button) shows:

• Timestamp, tool name, status badge, duration
• Per-row checkboxes with Select All / Deselect All / Delete Selected
• Filter by tool, status, and time range
• Click any row to see the parameter snapshot
• Replay Params — restores that run's parameters to the tool's form
• View Full Log — opens the full log in a resizable window with level filters and search
• DB size readout — the current on-disk size of the history database, next to the filters
• Clear History — drops every record, parameter, and log, then vacuums the database to reclaim the space. Application settings are preserved.

History is stored at ~/.decoui/history.db by default. Override with db_path in gui_main().

---

Example

See src/decoui/example/ for a complete demo covering every supported feature, in three groups: fields, running and form assist. Run it with python -m decoui.example.

Two tools there are worth reading as a pair before writing anything that runs for a while:

Tool	Shows
Demo Tools → Slow Task	progress() driving the bar, confirm=True, and what a failed run looks like. Cancellable on its own, because it only sleeps.
Demo Tools → Run Child Process	on_cancel around a real subprocess. Press Stop while it runs: the hook is what actually kills the child, and the with Popen(...) block is what reaps it.

Run it with:

uv run python main.py
# or
python main.py