Traceable Business Report (template-driven)

Render a multi-section analysis report from a template plus structured data, so every number in the report is resolved by the backend from a captured artifact result instead of being typed by the model. This is the default path for report generation: prefer it over hand-writing HTML.

Why

Hand-written report HTML is validated against the captured artifacts after the fact: a number the model remembers (or copies from an earlier report) does not exist in any artifact, the render is rejected, and repairing dozens of HTML numbers is slow. With this template the model only supplies pointers (artifact_id + result path); the backend resolves and formats the value, and a wrong pointer fails immediately with the exact selector to fix.

Tool call

{
  "template_path": "business-report/templates/business_report.html",
  "title": "Olist 巴西站经营复盘报告",
  "data": {
    "REPORT_SPEC": { "...": "see below" }
  }
}

data.REPORT_SPEC may be a JSON object or a JSON string. Do not pass html together with template_path. Prefer the object form: a spec hand-escaped into a JSON string is a frequent, fatal mistake — one wrong escape makes the whole call unparsable and nothing renders.

{
  "template_path": "business-report/templates/business_report.html",
  "title": "Olist 巴西站经营复盘报告",
  "data": {"REPORT_SPEC": {"title": "…"}}
}

Reference nodes (the only way to put a value in the report)

Every value is a reference node; the backend replaces it with the captured value and binds the rendered <span> to the producing artifact:

{"artifact_id": "sql-f6f8810...", "path": ["rows", 0, "gmv"], "format": "money2"}
{"artifact_id": "sql-29da2e3...", "path": ["rows", 0, "product_gmv"], "format": "money2"}
{"artifact_id": "sql-7de826c...", "kind": "sql", "quote": "r.score <= 2"}
{"artifact_id": "reference-e7ae093...", "format": "text"}

• field names a key of a Python artifact result (result_selector_examples in the manifest lists them). For a SQL metric use path ["rows", <index>, "<column>"] with a name from the manifest columns (the result is {"columns": [...], "rows": [...]}). A SQL selector that names a field is refused with RESULT_SELECTOR_INVALID, which lists the available keys.
• Python artifacts must expose their metrics as a JSON result field (for example {"RESULT_JSON": "{...}"}); reference {"artifact_id": "python-…", "field": "RESULT_JSON", "path": ["driver_analysis", "items", 0, "corr"]}.
• format: money2, num2, num1, int, pct (adds %), text/raw.
• Never inline a literal number. Prose numbers must also be reference nodes: a section whose text contains digits must carry evidence (a KPI ref, a table built from artifact_id + columns, a chart, or bound values in intro/bullets), otherwise the render is refused before the browser with REPORT_TEMPLATE_SPEC_INVALID, which names the section and lists the available artifacts.
• Layout labels (headings, table headers, section titles) must contain no digits; use reference nodes for anything numeric, including threshold labels.
• Get real artifact IDs and result fields from report_evidence_manifest before writing the spec.
• Start with artifact_types=["sql", "python", "chart"], then select the needed artifact_ids with detail=true (at most twelve per call). Read the governing KB slices separately. A catalog contains identifiers and columns, not the results; read a persisted output when the preview is truncated. A new report agent receives its own first catalog even if the lead already read the same version.
• Numeric preflight reports all unbound literals together, with spec JSON paths and SQL/Python producer candidates. Correct every listed value; adding one KPI to the section does not bind the remaining prose. Use a producer table plus a qualitative summary when per-value prose would be too long. Uncaptured derived totals require a captured calculation or a summary without that number.

REPORT_SPEC shape

{
  "title": "…", "subtitle": "…", "scope": "分析窗口 ｜ 口径 ｜ 货币",
  "kpis": [{"label": "商品 GMV", "unit": "R$", "hint": "SUM(price)", "ref": <ref>}],
  "sections": [
    {"heading": "一、总体经营表现", "kpis": [...], "intro": [[part, …]], "table": {...},
     "chart": {...}, "bullets": [[part, …]], "list": [[part, …]],
     "recommendations": [{"badge": "P0", "cls": "p0", "title": "…",
                          "lines": [{"label": "依据", "parts": [part, …]}]}],
     "note": {"label": "口径说明", "parts": [part, …], "ref": <ref>}}
  ]
}

• <part> is either a plain string (no digits) or a reference node.

• table: name the producer and the columns; the engine expands the rows. Never write one reference node per cell — a 10x5 table would cost ~5,500 characters and overrun the model output limit.

  {"artifact_id": "sql-f6f88101…", "label": "品类明细", "limit": 10,
   "columns": ["category", {"field": "gmv", "label": "GMV", "format": "money2"}, "orders_cnt"]}

  A columns entry is a field name (result key, also used as the label) or {"field", "label", "format"}. limit/offset window the artifact rows; each expanded cell carries its own path selector. The explicit "rows": [[part, …], …] form is only for cells that no artifact row can produce.

• chart: only when a real ChartArtifact exists. Image chart: {"refs": ["<sql-id>", "<chart-id>"], "img": "artifact://….json", "caption": "…"}. ECharts host (the engine then needs one producer with a selector and a target id): {"refs": ["<sql-id>", "<chart-id>"], "renderer": "echarts", "chart_id": "<chart-id>", "host_id": "chart-1", "producer": {"artifact_id": "<python-id>", "field": "CHART_DATA_JSON"}, "caption": "…"}. <figure>/<img> are chart hosts and must be covered by a chart artifact.
• recommendations produce <li> items; give at least as many as the user requested, and keep the section heading containing “建议”.
• note.ref: bind the knowledge-space slice actually used for the methodology statement. A whole-page binding does not count.

Chart configuration

Charts are not limited to the automatic line/bar/scatter choice. Use chart.option for a full JSON ECharts configuration, including any built-in series type, composed series, dataset/encode, multiple coordinates and zoom controls. For example:

{"caption":"Composition","producer":{"artifact_id":"sql-real","path":[]},
 "option":{"series":[{"type":"pie","radius":["40%","70%"],
 "encode":{"itemName":"actual_category","value":"actual_measure"}}]}}

The engine injects the producer's captured rows/columns as the default dataset. For structured charts, select series.data, nodes, links, dataset.source, or individual data values using {artifact_id,field,path} references. Do not inline observed data. Use captured tree/graph/boxplot/candlestick structures; request only a missing conversion via request_report_calculation when necessary. Map charts can set maps={mapName:{artifact_id,field,path}} to register captured GeoJSON, then use that mapName in the option. All input IDs, selectors, resolved options and geography are captured and checked against the rendered chart. No chart-type whitelist is imposed; JavaScript callbacks and extra plugins are not executable through JSON options. Captured Python chart images remain supported. Placeholder-based skill templates use data.REPORT_CHARTS=[<chart>, ...] for the same renderer, independently of their preset chart layout.

Output budget

The whole spec travels in one tool-call argument, and the model output limit is about 8k tokens: a payload that is cut off cannot be parsed, so the render never even starts. Keep the spec small — tables by artifact_id + columns (never per-cell references), limit instead of pasted rows, short intro/bullets, and no repeated prose. If a render result says the arguments were cut off by the output limit, shrink the spec instead of resending it.

A long report may be built in pieces. Send the first part as data={"REPORT_SPEC": {...}}, then extend the saved spec with data={"REPORT_SPEC_APPEND": {...}} (call it again with retry_saved=true when a render was already attempted): sections and kpis are appended, every other key overwrites. The engine stores the merged spec, so never resend a part you already sent, and finish with the required recommendation section.

KPI values must use ref, not value. Paragraphs use sections[].intro as arrays of parts; recommendations belong to sections[].recommendations and contain a title and lines[].parts. Unknown keys such as placeholders, sections[].parts or top-level recommendations are rejected with a JSON path before rendering. REPORT_SPEC repairs retain unspecified root fields; an explicitly supplied sections or kpis replaces that complete list. A root field set to null is removed, including an unsupported field from a previous invalid spec. REPORT_SPEC_APPEND extends those lists and does not insert an identical fragment twice. Preflight (6), document assembly (6), actual render validation (3), and invalid calls (3) have separate budgets and a total cap of 18 report calls. Pre-render spec/selector failures never spend render attempts. Progress compares the merged spec and error; retry_saved and template HTML length are not progress signals. Fix all returned JSON paths together. An identical input with the same failure stops retries.

The report agent's tool schema is generated from this same spec model. Captured charts belong only in sections[].chart: {"refs":["real-chart-id"],"caption":"…"}. Image URLs are resolved from the captured chart output. Native figures use the existing AutoChart frontend. ECharts still uses its existing reviewed template and bootstrap; an empty div without rendering code is not a chart.

If an essential derived result is truly missing, call request_report_calculation with its name, formula and real SQL/Python source_artifact_ids, then terminate. The lead computes only that missing result with code_interpreter using the injected REPORT_CALCULATION_INPUTS JSON file and prints JSON for normal artifact capture. Resume the saved spec and bind the new Python artifact; never rerun full analysis.

For an authored recommendation target, a reference node in recommendation lines may add "proposal":"reduce the rate to 5%". The engine labels it explicitly as a proposed, unobserved target. The node's artifact_id and field/path still select the real rationale; proposal is forbidden in KPI and observed-value paragraphs.

On failure, a labelled draft can preview independently validated blocks through the existing HTML preview. It does not create a report_id or indicate completion. Use status=completed to determine success, never the mere presence of draft HTML.

Failure handling

For a large nested result, request only its relevant subtree with report_evidence_manifest(artifact_ids=["real-id"], detail=true, selectors={"real-id":{"field":"CHART_DATA_JSON","path":["stats_table"]}}). Selector examples also name fields inside captured JSON strings; truncated previews are not evidence that a result is missing. This reuses the manifest and result selector, and never reexecutes analysis.

Nested Python tables may use the same field/path selector on the table node, for example {"artifact_id":"python-id","field":"CHART_DATA_JSON", "path":["stats_table"],"columns":["actual_column"]}. Prefer an object with rows and columns. The engine expands both SQL and Python rows and retains the complete selector for every cell; do not paste per-cell values into the spec.

The render error names the exact artifact_id, selector and container. Fix the spec (data.REPORT_SPEC), then call retry_saved=true with the corrected data only — never resend full HTML and never rerun completed analysis.