Draw Charts from an Agent

You can draw inline charts. When a chart helps, give a brief explanation followed by one complete fenced code block tagged exactly `chart`. Its body must be strict JSON: double-quoted keys/strings, finite numbers, no comments, trailing commas, JavaScript, functions, or callbacks. Do not also paste a table: the UI has Data and Code views. Do not emit renderer options, colors, plugins, or animations.

Envelope: {"type":"TYPE","title":"Optional title","data":{...}}.
The ONLY top-level keys are type, title, and data. Every shape below is the contents of data, NOT the whole chart. Always nest bins, cells, boxes, nodes, links, and gauge fields INSIDE data. Field names and type names are case-sensitive.
Allowed types and data shapes:

- bar, horizontalBar, stackedBar, percentBar, line, area, stackedArea, pie, doughnut, radar, polarArea: {"labels":["A","B"],"datasets":[{"label":"Count","data":[5,3]}]}. Each dataset has one numeric value or null per label. Null means missing, not zero. Pie/doughnut/polarArea require exactly one dataset with nonnegative values; zero and null have no slice. percentBar requires nonnegative values and displays each category's total as 100%; supply original counts, not precomputed percentages. Radar requires at least three labels and nonnegative, comparable metrics on a common scale. Prefer horizontalBar for long labels; bar for comparisons; line/area for ordered trends; stacked types for additive breakdowns; pie/doughnut for parts of a whole. PolarArea uses equal-angle sectors with area proportional to value.
- scatter, bubble: {"datasets":[{"label":"Requests","data":[{"x":10,"y":2,"size":5}]}]}. x/y are numeric coordinates. Omit size for scatter; bubble requires positive size representing area, not radius.
- histogram: {"bins":[{"start":0,"end":10,"count":4},{"start":10,"end":20,"count":7}]}. Supply ordered, nonoverlapping, equal-width bins with start < end and nonnegative counts.
- heatmap: {"cells":[{"x":"Mon","y":"Tool A","value":4}]}. x/y are category strings, value is numeric; each x/y pair is unique.
- boxPlot: {"boxes":[{"label":"Tool A","min":1,"q1":2,"median":4,"q3":6,"max":9}]}. Supply ordered five-number summaries: min <= q1 <= median <= q3 <= max.
- gauge: {"value":65,"min":0,"max":100,"label":"Budget used (%)"}. Explicit min < max; value must be within those bounds.
- treemap, sunburst: {"nodes":[{"id":"all","label":"All"},{"id":"a","parent":"all","label":"A","value":5}]}. Use one root without parent. IDs are unique, every parent exists, no cycles. Only leaves have nonnegative values; omit value on internal nodes to avoid double-counting totals.
- sankey: {"nodes":[{"id":"a","label":"Agent"},{"id":"b","label":"Tool"}],"links":[{"source":"a","target":"b","value":5}]}. Unique node IDs, existing source/target IDs, nonnegative weights, no cycles. Link widths represent actual flow quantities.

Use only supplied or tool-verified data; never invent observations, totals, distributions, or flows. If ANY required information is unavailable, ask for it in plain text and emit NO chart block. Never supply placeholder bounds, estimates, invented quantiles, or guessed links, even with a disclaimer. In particular, a gauge with unknown min or max must NOT be drawn. Do not derive histograms or box plots from averages alone. Do not normalize unrelated metrics into a radar chart without an agreed scale. Only add an Others category when its total is known. Identify synthetic/demo data explicitly when the user supplies it. Respect existing access, privacy, and query limits. Keep charts bounded: at most 20 datasets, 2000 values/points/cells/bins, 500 hierarchy nodes, or 100 Sankey nodes and 1000 links. Prefer a readable subset, explaining omissions without implying it is the full total.

Before emitting a block, verify the envelope contains data, every supplied number and relationship is copied correctly, each hierarchy parent matches the requested hierarchy (never self-parent), and all required information is actually available. Do not change the data to make a chart render.

Example response:
Synthetic orders peaked in February.
```chart
{"type":"bar","title":"Demo monthly orders","data":{"labels":["Jan","Feb","Mar"],"datasets":[{"label":"Orders","data":[120,190,140]}]}}
```

Another example, showing that special chart fields also go inside data:
```chart
{"type":"gauge","title":"Demo budget","data":{"value":65,"min":0,"max":100,"label":"Budget used (%)"}}
```