analyze-xlsx

name	analyze-xlsx
description	Analyze an ingested .xlsx document for formula errors, suspected typos, outliers, missing totals, and dead references
model	deepseek/deepseek-v4-flash
tier	flash
inputs	[{"name":"document_id","type":"str","required":true,"description":"Prisma Document.id of an .xlsx that was ingested via Phase 33"},{"name":"max_issues","type":"int","required":false,"description":"Soft cap on number of issues returned (default 20)"}]
outputs	[{"name":"document_id","type":"str"},{"name":"title","type":"str"},{"name":"sheet_count","type":"int"},{"name":"issue_count","type":"int"},{"name":"issues","type":"list[dict]","description":"[{severity, category, sheet, cell, description, suggestion}]"},{"name":"model","type":"str"}]

analyze-xlsx

Audits a Phase-33-ingested Excel workbook for data quality problems. The skill walks every page (one page = one worksheet) of the ingested document, collects the Markdown/text representation produced during ingestion, and asks DeepSeek to identify problems across six categories.

How it works

1. Fetches the document record from the database using document_id.
2. Verifies the document was ingested as an Excel file (sourceFormat is xlsx or xls). Rejects other formats with a 400 error.
3. Walks the chapter → section → page hierarchy produced by the Phase-33 ingestion pipeline (falls back to a flat page list for older shapes).
4. Caps each page's text at 4 000 characters to stay within model context limits, then assembles all sheets into one prompt blob separated by --- next sheet ---.
5. Sends the assembled prompt to DeepSeek (model configured via OFFICEPLANE_AGENT_MODEL_FLASH, default deepseek/deepseek-v4-flash) with temperature=0 and response_format={"type": "json_object"} for deterministic, machine-readable output.
6. Normalises and validates every issue returned by the model:
   • Unknown severity values are coerced to "low".
   • Issues with an unrecognised category are silently dropped.
   • The list is hard-capped at max_issues (default 20, max 100).
7. Writes a SkillInvocation audit row via persist_skill_invocation.

Issue categories

Category	What it flags
formula_error	Formula returns #REF!, #DIV/0!, #VALUE!, or similar Excel errors
suspected_typo	A number that looks off-by-magnitude relative to its column peers
outlier	A value dramatically larger or smaller than its neighbours
missing_total	A column of numeric values with no totals row when one is expected
dead_reference	A formula references a cell or sheet name that does not exist
inconsistent_format	Mixed text/numbers in a column, or inconsistent units (%, decimal, $)

Severity levels

• high — likely causes incorrect calculations or broken reports
• medium — suspicious but may be intentional; warrants review
• low — cosmetic or style issue

Output shape

{
  "document_id": "cuid...",
  "title": "Q3 Budget Review",
  "sheet_count": 4,
  "issue_count": 3,
  "issues": [
    {
      "severity": "high",
      "category": "formula_error",
      "sheet": "Summary",
      "cell": "D12",
      "description": "=SUM(B2:B11) returns #REF! — referenced range is outside the sheet",
      "suggestion": "Check that the source range B2:B11 exists and extend it if needed"
    }
  ],
  "model": "deepseek/deepseek-v4-flash"
}

If no issues are found the issues list is empty and issue_count is 0. If the workbook has no page content (empty ingestion), the skill returns early with sheet_count=0 and a note field explaining the situation.