toolr-command-authoring - SKILL.md Agent Skill

name: toolr-command-authoring description: | Author toolr commands in a project's own `tools/*.py` files. Use when adding, editing, or refactoring a toolr command, group, or context hook; when introducing a new `tools/` directory; when wiring `@command`, `@command_group`, `@arg`, or `@arg_section` decorators; when configuring a command's docstring-driven `--help`; or when debugging "command not found" / "manifest stale" errors against toolr. Triggers on phrases like "add a toolr command", "extend toolr", "wire a new toolr group", "toolr tools/", `@command_group`, `ctx.run`, `toolr-manifest.json`. Stays inert in projects that don't use toolr and on requests to package commands as a distributable plugin (covered by the `toolr-command-packaging` skill).

Authoring toolr commands

You are extending an existing toolr project — a repo that already has (or will have) a tools/ directory at the root and a toolr binary on PATH. Your job is to add or change tools/*.py files so the user gets new subcommands under toolr ....

This skill teaches the shape of authoring toolr commands. For the actual surface of decorators, Context methods, and docstring conventions, consult the generated references in references/ — they are rebuilt from toolr's own source on every release, so they cannot drift.

Workflow

Confirm the project is a toolr project. Look for tools/ at the repo root with tools/pyproject.toml. If it doesn't exist, ask the user to run toolr project init first — that scaffolds everything correctly and you should not duplicate it.
Decide where the command goes. One file per subcommand is the easy default; group multiple commands in the same file when they share helpers or a parent group.
Declare (or import) a group with command_group(...). The group is what toolr <group> <cmd> selects on. Reuse an existing group across files by passing group="..." to @command rather than redeclaring it.
Write the function. Take ctx: Context as the first parameter, then your CLI arguments as ordinary Python parameters. Type hints drive argparse binding; defaults make arguments optional; Annotated[T, arg(...)] adds clap metadata (aliases, metavar, help_section, must_exist, etc.).
Document via Google-style docstring. The first line is the short help (toolr <group> --help). The rest is the long help (toolr <group> <cmd> --help). Args: populates per-argument help.
Try it. toolr <group> <cmd> --help builds the manifest on the fly if it's stale (the freshness work landed in 0.20.0); on older toolr fall back to toolr project manifest rebuild. If the command doesn't appear, the manifest builder rejected it — read the error.

What "looks right" looks like

A canonical single-file command:

"""Long-form group description, used as the group's `--help` long text."""

from toolr import Context, arg, command, command_group

command_group("greet", "Say hello in various ways", docstring=__doc__)


@command(group="greet")
def hello(ctx: Context, who: str = "world", *, loud: bool = False) -> None:
    """Print a greeting.

    Args:
        who: Who to greet. Defaults to ``"world"``.
        loud: Shout instead of speak.
    """
    msg = f"Hello, {who}!"
    if loud:
        msg = msg.upper()
    ctx.info(msg)

Add this to tools/greet.py, then toolr greet hello --help works.

Common authoring moves

Group across files. Pass group="ci.build" to @command from any file; only one file needs to call command_group("ci.build", ...).
Nested subgroups. A dotted name (command_group("docker.image", ...)) attaches under the parent named before the last dot.
Help sections. Build an ArgSection at module scope via arg_section("Logging", description="..."), then attach it via Annotated[bool, arg(help_section=LOGGING)]. Same ArgSection object across all members or you'll silently create duplicate sections.
Calling subprocesses. Use ctx.run(...); it inherits stderr for TTY-aware tools and propagates timeouts.

Runtime working directory

Commands run with the working directory set to the repo root, regardless of where you invoke toolr from (the make/cargo convention). Two consequences:

Relative path arguments resolve from the repo root, not your current directory. toolr build ./out.txt run from tools/ writes <repo-root>/out.txt, not <repo-root>/tools/out.txt. Pass an absolute path when you need a file relative to where you ran the command. toolr prints a one-line note on stderr if you pass a relative path argument from a subdirectory, so the surprise is visible.
ctx.run(...) subprocesses inherit the repo root as their cwd unless you override it, so paths in the commands you spawn are also repo-root-relative by default.

Static-only discovery contract

toolr discovers commands only by static analysis of tools/*.py — it never imports or executes your modules to build the manifest. Declare command_group(...) at module top level and apply @command / @group.command to module-level functions. Commands registered dynamically — in a for loop, behind an if, or returned from a factory called at import time — are not discovered and will not appear in --help, completion, or dispatch. If a command is missing, make its registration a top-level, statically-visible declaration.

Anti-patterns

Don't register commands dynamically. A loop like for name in names: group.command(...) or a factory that builds commands at import time produces nothing — the static parser can't see it. Declare each command at module level (see the static-only contract above).
Don't bypass the decorator surface. Defining a function and then calling register(...) directly skips the manifest builder. Always @command or @<group>.command.
Don't reach into toolr._* internals. Those modules are implementation detail; the public surface is exactly the names in from toolr import (...), which references/commands.md lists.
Don't pass description= and docstring= to the same command_group(...). It raises. Pick one — docstring=__doc__ is the canonical form when the module's docstring is the long description.
Don't write your own argparse subparser. toolr owns the parser; you describe shape via decorators and type hints.

References

references/commands.md — every name exposed by import toolr. Signatures, defaults, annotations, and docstrings, regenerated from toolr.__all__ on every release. Treat it as the source of truth for the decorator API.
references/docstrings.md — exactly which Google-style section headers toolr's docstring parser recognises and how each is rendered. Generated from the same KNOWN_SECTION_HEADERS table the parser reads at runtime.

Local feedback loop

toolr <group> --help — list commands in the group; if yours is missing, the manifest builder rejected it.
toolr <group> <cmd> --help — full per-command help.
toolr project manifest rebuild --force — bypass freshness detection and rebuild from scratch.
On a manifest error, the message points at the offending line in tools/*.py. Fix the source; the manifest auto-rebuilds on the next dispatch.

Packaging is a different problem

If the user wants to ship an existing set of toolr commands as a distributable Python plugin (so other projects can pip install and get the commands), that is the toolr-command-packaging skill's job. This skill does not cover wheel-building, manifest embedding, or PyPI publishing — invoke the packaging skill for that work.

CI is a different problem

If the user wants to run these commands in GitHub Actions (a caller workflow that installs toolr, sets up the venv, and runs toolr <group> <cmd>), that is the toolr-ci-setup skill's job. This skill does not cover the s0undt3ch/ToolR action, pinning policy, or CI cache shapes — invoke the CI-setup skill for that work.