data-analysis

Data Analysis Skill

Answers questions about a bundled sales dataset (resources/sales.csv): 96 rows covering six months, four regions, four product categories and three sales channels.

This is the reference skill for exercising sandbox executors. Both scripts are standard library only and read their data from the skill folder, so they do real work with no network, no third-party packages and no host access — exactly the shape a sandbox is meant to run.

When to Use This Skill

"Which region had the highest revenue?" → aggregate
"What were average units sold per category?" → aggregate
"How many rows are in the dataset and what does revenue look like?" → profile_dataset
"Compare online versus retail sales in the north" → aggregate with filters

Skill Scripts

profile_dataset

Reports dataset shape, column names, summary statistics for every numeric column (count, sum, mean, median, stdev, min, max) and distinct-value counts for the categorical ones.

column (optional): Profile a single numeric column instead of all of them

Exits 2 for an unknown column.

aggregate

Groups rows and aggregates a numeric column, optionally filtered and truncated.

group-by (required): Column to group by — month, region, category, channel
metric (optional): Numeric column to aggregate, default revenue
agg (optional): sum (default), mean, median, min, max, count
where (optional, repeatable): Filter as column=value, e.g. region=north
top (optional): Keep only the highest N groups

Exits 2 for an unknown column or malformed filter, 1 when no rows match.

Usage Examples

Revenue by region, highest first:

group-by: region

Top 3 categories by units sold, online only:

group-by: category
metric: units
agg: sum
where: channel=online
top: 3

Average revenue per month in the west:

group-by: month
agg: mean
where: region=west

Data Source

resources/sales.csv is generated sample data, included so the skill is self-contained and deterministic. It is not real sales data.

Sandbox Behaviour

Both scripts produce byte-identical output on the host and under LocalSandboxScriptExecutor or OpenSandboxScriptExecutor, including the ../resources/sales.csv read through the skill root and non-zero exit codes for bad arguments. That equivalence is the point: swapping the executor changes where the script runs, not what the model sees.