Contribution heatmap

The GitHub profile grid, rebuilt from its own raw material — and the subject is this repository. Every cell is one UTC day of pond's default branch; colour carries the non-merge commit count. A heat map is the right shape whenever the question is "when was the activity?" asked across two time scales at once: weeks along the x axis, the shape of a week down the y.

The chart

Point at a cell for its day and count. Sundays are the top row, Saturdays the bottom — GitHub's reading — and the columns run from the repo's first week (12 April 2026) to the current, deliberately partial one. The brightest cell is Wednesday 13 May: 35 commits; the heaviest column is the week of 2 August: 129.

The data

Real measured data — this repo's own. The GitHub REST API's commit_activity statistics for pond-ts/pond: the numbers behind the repository's own activity graph, retrieved the day this card was built. 1,001 commits across 126 days — 18 Sunday-start UTC weeks, from first commit to the moment of retrieval.

Column	Meaning
sun … sat	Non-merge commits that weekday, one row per UTC week

Three binning rules matter, and the generator refuses to emit if any stops holding. Merge commits are excluded — the API total matches git rev-list --count --no-merges origin/main exactly, which is how the claim was pinned. Days are UTC, so a late CEST evening's commits land in the next row. And the API serves a fixed 52-week window; the 34 leading all-zero weeks (the repo is four months old) are trimmed, while the current week keeps its honest trailing zeros — days that haven't happened yet.

What the grid shows about how pond is built: 26 of the 126 days are quiet; the other hundred average 10 commits a day, with a 26-day unbroken streak in there. And the usual profile-heatmap tell is missing — there is no weekend stripe. Sundays (142 commits) outwork Thursdays (113), because much of this history is multi-agent experiment sessions, and agents don't know it's Saturday.

Build it

1. Shape the data wide — and key the weeks as spans. <HeatMap> wants its y dimension as columns: one column per row of the grid, one sample per x bin — the same wide shape the Niño 3.4 grid uses for its 45 year-rows. The x bins here are range-keyed (timeRange, each week carrying its own start/end), because a week is a span: a point-keyed series would leave the layer to infer each cell's bin from the midpoints to its neighbours, and every cell would straddle its week's start by ±3.5 days. Same rule as bars on auction hours in negative prices.

const series = commitActivity(); // one row per week: sun, mon, … sat

<ChartContainer range={commitActivityRange()} width={width}>
  <ChartRow height={210}>
    <YAxis id="dow" ticks={COMMIT_ROW_TICKS} label="" width={44} />
    <Layers>
      <HeatMap series={series} columns={COMMIT_ROWS} colors={ramp} axis="dow" />
    </Layers>
  </ChartRow>
</ChartContainer>;

2. Order the rows for the reading you want. <HeatMap> draws the first column at the bottom, and the GitHub convention puts Sunday on top — so the row list runs sat → sun. The y ticks label only Mon / Wed / Fri (GitHub's three), each at its row's centre:

export const COMMIT_ROWS = ['sat', 'fri', 'thu', 'wed', 'tue', 'mon', 'sun'];
export const COMMIT_ROW_TICKS = [
  { at: 1.5, label: 'Fri' },
  { at: 3.5, label: 'Wed' },
  { at: 5.5, label: 'Mon' },
];

3. Read cells with a real 2-D hit. The hover readout is fed by <Selector onHover>, which resolves the cell under the pointer on both axes. The 1-D tracker can't do this — it samples every row at the cursor's x and knows nothing about y:

<Selector onHover={setHit}>
  <ChartRow height={210}>…</ChartRow>
</Selector>

The hit's label is the row (the weekday column's name), its key is the week's start — together they name one day, which is what the readout prints.

Options to try

Option	What it does	Reach for it when
colors	The ramp the counts map into	This card uses the site's sequential ramp; a fixed-hue ramp survives theme changes
domain	Pins the colour scale's extent	Comparing two heat maps, or keeping colours stable as data updates — see Niño 3.4
Re-bin x with pond	aggregate(Sequence.calendar('month'))	Months instead of weeks — climate-stripes-style summaries of the same columns
Fewer / more ticks	Which rows get named	Seven rows read fine with three labels; forty-five need explicit picks
timeFormat	The x axis's date rendering	Weekly bins often want month names, not day-precision dates

See also

• Measles and the vaccine — the same layer at 50 rows × 81 years, with a log-scaled colour domain
• Niño 3.4 — the wide-series shape this grid borrows, and the re-binning story
• Storybook — the systematic knob walk