| name | reporting |
| description | Synthesize the profile, analysis tables, charts, and any model outputs into a single structured report.json — the final deliverable consumed by the web app. |
Reporting
The final step. It gathers everything you produced (data/profile.json, data/analysis/*.csv, the PNGs in charts/, and any model outputs) and uses Gemini to synthesize a decision-ready business report at data/report.json.
Prerequisites
Before running, make sure you have:
- Run the
data-explorer skill (data/profile.json exists).
- Saved your answer tables to
data/analysis/*.csv.
- Rendered the supporting charts to
charts/*.png.
Script
python3 /.agents/skills/reporting/scripts/build_report.py \
--workspace ./workspace \
--question "Which categories and customers drive the most revenue?" \
--dataset-name "Northwind Orders"
| Argument | Required | Description |
|---|
--workspace | yes | Root workspace directory |
--question | yes | The user's original business question |
--dataset-name | no | Friendly dataset name (defaults to "Dataset") |
--charts | no | Comma-separated file:title pairs to force chart order/captions; otherwise all PNGs in charts/ are auto-included |
What it does
- Reads
data/profile.json and every CSV in data/analysis/.
- Lists the PNGs in
charts/ (these become the charts[] entries).
- Sends the question + profile summary + analysis tables to Gemini via the Interactions API and asks for a structured business report.
- Validates the JSON against the output contract and writes
data/report.json. If the model output is unusable, it writes a deterministic fallback report built directly from the analysis tables — so a report.json is ALWAYS produced.
Output — data/report.json
See the Output Contract in AGENTS.md. Key points:
executive_summary must answer the question directly in 2-4 sentences.insights[] are ranked, each with a concrete metric/value pulled from the analysis tables.charts[].file must point to a real PNG under charts/.tables[] are small (≤ 20 rows) display tables drawn from data/analysis/*.csv.recommendations[] are concrete and actionable.
Rules
- Never fabricate numbers — every figure must trace back to the profile or an analysis table.
- If a needed input is missing, still emit a valid
report.json describing what could and could not be answered.
