probabl-ai/skore-skills/add-python-package

Add Python Package

The only skill that knows python -m skore_skills env add. Callers must not splice manager commands themselves.

Human-facing prose

Details: setup-workspace references/human_facing_prose.md. Run python -m skore_skills … yourself. In questions and replies, show the manager command from print-only stdout (pixi add …, uv add …, pip install …) or a plain-language intent. Never paste python -m skore_skills, env add, env add-skore, or --feature / --group as something the user should run or choose. Never paste env graphviz either; quote that command's instructions or printed manager line.

Pre-flight

Tick, then immediately run the matching sequence step. Do not stop after listing the boxes.

- [ ] status + env detect
- [ ] env.managed: null and setup-ml-project installed → load it and stop | null → say unresolved and stop | false and setup already chose unmanaged → name package and manager, return | false → ask, no --execute | true → continue
- [ ] classify: editable | add-skore | env route then env add
- [ ] skrub → env graphviz (install conda Graphviz and verify SVG when allowed)

Sequence

1. python -m skore_skills status and env detect.
2. If policy.env.managed is null: env is unresolved. If

status.skills.setup-ml-project is true, load setup-ml-project and stop. Do not ask whether to set up an environment. Do not load setup-python-env. If that skill is not installed, say the environment is unresolved and stop. Do not bootstrap here.

1. Classify this turn before adding anything:

• Editable only if the user asked to install the workspace

package, or setup-ml-project selected that box. has_src true is not enough.

• Skore if the package is Skore.
• Named package otherwise (skrub, pandas, …). Never

--editable for a named dependency.

1. If managed is false and setup-ml-project dispatched this

turn: the opening question already chose user-managed. Name the package and the manager in words, then return. Do not run env add — print-only refuses while managed is false. Do not invent pixi add / uv add / pip install. Do not ask. Do not wait. Do not say the turn continues after the user confirms. Stop this sequence. Do not continue at step 5 or step 6.

1. If managed is false on a standalone turn: do not

--execute. Run print-only env add (or env add --editable, or env add-skore --mode <mode>) to obtain the manager line. Every ask in this skill carries its context inline: name the package(s), the manager and env the command would touch, and what each option does. Do not name the calling skill. A file link is an addition, never the context. Ask with two options:

1. I will handle it (default) — name the package(s) and

show that stdout (e.g. pixi add pandas). Do not wait; return.

1. Please install this now — show the same manager line,

wait until the user confirms it is done, then return.

Show exactly one manager command. Do not list agent extras, optional extras, ask, or refuse unless env route JSON this turn returned that scope. If print-only did not run, name the package and the manager in words. Do not invent a wrapper command.

1. If this turn is editable and has_src is true:

``bash python -m skore_skills env add --editable --execute ``

Never pip install -e . in a pixi project. If has_src is false, stop. Name setup-workspace only if it is installed. A non-zero exit is not a successful install. Do not describe it as done.

1. If the package is Skore, read the persisted policy.skore_mode

and run:

``bash python -m skore_skills env add-skore --mode <mode> --execute ``

If the mode is unset, return to evaluate-ml-pipeline. Do not spell skore[...] or pick conda vs PyPI yourself. See add-python-package/references/skore_variant.md.

1. Else python -m skore_skills env route <pkg>. Use only the

scope returned this turn. Do not list other branches. Do not guess default.

• No JSON this turn → name env route <pkg> and stop. Do not

env add --execute.

• refuse → stop. Quote message. Do not install.
• default → env add --execute <pkg>
• agent → env add --feature agent --execute <pkg>
• ask → One question, then stop. Two options only: the

project runtime, or a named optional extra / agent tools. Do not add a context line, a manager command, or a flag to either option. Do not preview pixi add. Record only python -m skore_skills env route <pkg>. The install command waits until the next turn, after they answer.

When managed is true and scope is not refuse, pass --execute on that one command. Never paste pixi add / uv add / pip install from memory.

1. If this turn is skrub, Graphviz is required for DataOp HTML

graphs. An already-importable skrub still runs this step. Do not return before it because import skrub succeeds. After the add, run python -m skore_skills env graphviz. Then:

• dot is set, or action is conda and managed →

python -m skore_skills env graphviz --execute (installs conda Graphviz when needed, then verifies pydot can render SVG in the composed env).

• action is system and dot is null → AskUserQuestion

with two options, quoting JSON instructions only:

1. I will install Graphviz (default) — do not wait;

return.

1. Please wait until I confirm — wait, then re-run

env graphviz --execute to verify SVG rendering.

• Unmanaged → do not --execute. Show command (conda) or

instructions (system) from print-only JSON.

If --execute fails, quote its repair diagnostic and end this install sequence. Return that diagnostic to the caller. Do not discard analysis the caller already finished. When this turn's only job was adding the package, that return ends the turn. Never prescribe dot -c, admin rights, or a package-manager repair from memory. Never invent brew / apt / dnf / pacman / zypper / winget from memory. Do not pip-install Graphviz or add it as a Python package on uv / poetry / hatch / pip-venv. Refuse that in prose; do not paste env add graphviz or uv add graphviz as a command fence.

Return when the import is available, when the user confirmed they installed it, when they chose to handle it themselves, or when this setup turn already chose user-managed and the package was named. That last return does not wait.

References

• references/skore_variant.md