Data Tables Guide

This guide covers how to present structured data using datatable and spreadsheet blocks, and how to use the transform_data tool for large datasets.

Overview

Fabric Agents supports three ways to display tabular data:

Key principle: For datasets with 20+ rows, use the transform_data tool to write data to a JSON file and reference it via "src" instead of inlining all rows. This dramatically reduces token usage and cost.

Inline Tables (Small Datasets)

For datasets under 20 rows, inline the data directly in the markdown block:

Datatable

```datatable
{
  "title": "Top Users",
  "columns": [
    { "key": "name", "label": "Name", "type": "text" },
    { "key": "revenue", "label": "Revenue", "type": "currency" },
    { "key": "growth", "label": "Growth", "type": "percent" },
    { "key": "active", "label": "Active", "type": "boolean" },
    { "key": "tier", "label": "Tier", "type": "badge" }
  ],
  "rows": [
    { "name": "Acme Corp", "revenue": 4200000, "growth": 0.152, "active": true, "tier": "Enterprise" },
    { "name": "StartupCo", "revenue": 85000, "growth": -0.03, "active": true, "tier": "Starter" }
  ]
}
```

Spreadsheet

```spreadsheet
{
  "filename": "q4-revenue.xlsx",
  "sheetName": "Revenue",
  "columns": [
    { "key": "month", "label": "Month", "type": "text" },
    { "key": "revenue", "label": "Revenue", "type": "currency" }
  ],
  "rows": [
    { "month": "October", "revenue": 125000 },
    { "month": "November", "revenue": 142000 }
  ]
}
```

Column Types Reference

Important notes:

• currency — Pass the raw number, NOT a formatted string. 4200000 renders as $4,200,000.
• percent — Pass as decimal. 0.152 renders as +15.2%. Positive values are green, negative are red.
• boolean — Use actual true/false, not strings.

File-Backed Tables (Large Datasets)

When to Use

Use the transform_data tool + "src" field when:

• Dataset has 20+ rows — inlining costs ~$1+ in tokens for 100 rows
• Data comes from a large API response or tool result
• You need to filter, reshape, or aggregate raw data before display
• Data is in CSV, TSV, or unstructured text that needs parsing
• You want to join data from multiple sources

The transform_data Tool

transform_data runs a script in an isolated subprocess that reads input files and writes structured JSON output.

Parameters:

Path conventions:

• Input files are relative to the session directory. Common locations:
  • long_responses/tool_result_abc.txt — saved tool results
  • data/previous_output.json — output from a prior transform
  • attachments/data.csv — user-attached files
• Output file is relative to the session data/ directory. Just provide the filename (e.g., "transactions.json")

Script argument conventions:

• Input file paths are passed as positional command-line arguments
• The last argument is always the output file path
• Python: sys.argv[1:-1] = input files, sys.argv[-1] = output path
• Node/Bun: process.argv.slice(2, -1) = input files, process.argv.at(-1) = output path

Output JSON Schema

The output file should contain valid JSON in one of these formats:

Full format (recommended):

{
  "title": "Recent Transactions",
  "columns": [
    { "key": "date", "label": "Date", "type": "date" },
    { "key": "amount", "label": "Amount", "type": "currency" },
    { "key": "status", "label": "Status", "type": "badge" }
  ],
  "rows": [
    { "date": "2025-01-15", "amount": 250.00, "status": "Completed" }
  ]
}

Rows-only format:

{
  "rows": [
    { "date": "2025-01-15", "amount": 250.00, "status": "Completed" }
  ]
}

Or just a bare array:

[
  { "date": "2025-01-15", "amount": 250.00, "status": "Completed" }
]

Merge semantics: When using "src", inline columns and title in the markdown block take precedence over values in the file. This lets you define column types in the block while pulling rows from the file.

Referencing the Output

After transform_data succeeds, it returns the absolute path to the output file. Use that exact path as the "src" value in your datatable or spreadsheet block:

```datatable
{
  "src": "/absolute/path/returned/by/transform_data",
  "title": "Recent Transactions",
  "columns": [
    { "key": "date", "label": "Date", "type": "date" },
    { "key": "amount", "label": "Amount", "type": "currency" },
    { "key": "status", "label": "Status", "type": "badge" }
  ]
}
```

Important: Always use the absolute path from the transform_data tool result. Do not construct relative paths manually.

Complete Workflow Example

User asks: "Show me all Stripe transactions from last month"

Step 1: Call the Stripe API via MCP tool — get large JSON response

Step 2: Call transform_data to extract and structure the data:

transform_data({
  language: "python3",
  script: "import json, sys\nwith open(sys.argv[1]) as f:\n    data = json.load(f)\nrows = [{\n    'id': t['id'],\n    'date': t['created'],\n    'amount': t['amount'] / 100,\n    'status': t['status'].title(),\n    'customer': t.get('customer_email', 'N/A')\n} for t in data.get('data', data.get('transactions', []))]\nwith open(sys.argv[-1], 'w') as f:\n    json.dump({'rows': rows}, f)",
  inputFiles: ["long_responses/stripe_result.txt"],
  outputFile: "transactions.json"
})

Step 3: Output the datatable block using the absolute path from transform_data result:

```datatable
{
  "src": "/absolute/path/from/transform_data/result",
  "title": "Stripe Transactions — Last Month",
  "columns": [
    { "key": "id", "label": "ID", "type": "text" },
    { "key": "date", "label": "Date", "type": "date" },
    { "key": "amount", "label": "Amount", "type": "currency" },
    { "key": "status", "label": "Status", "type": "badge" },
    { "key": "customer", "label": "Customer", "type": "text" }
  ]
}
```

Common Patterns & Recipes

JSON API Response → Datatable

Most common pattern. Extract fields from a JSON API response:

Python:

import json, sys

with open(sys.argv[1]) as f:
    data = json.load(f)

# Handle common API response shapes
items = data.get('data', data.get('items', data.get('results', data)))
if not isinstance(items, list):
    items = [items]

rows = [{
    'id': item['id'],
    'name': item.get('name', ''),
    'created': item.get('created_at', ''),
} for item in items]

with open(sys.argv[-1], 'w') as f:
    json.dump({'rows': rows}, f)

CSV/TSV → Spreadsheet

Parse CSV data into a spreadsheet for export:

Python:

import csv, json, sys

with open(sys.argv[1]) as f:
    reader = csv.DictReader(f)
    rows = list(reader)

# Auto-detect columns from CSV headers
columns = [{'key': k, 'label': k.replace('_', ' ').title(), 'type': 'text'} for k in rows[0].keys()] if rows else []

with open(sys.argv[-1], 'w') as f:
    json.dump({'columns': columns, 'rows': rows}, f)

Multi-Source Join

Combine data from multiple tool results:

Python:

import json, sys

# sys.argv[1:-1] are input files, sys.argv[-1] is output
with open(sys.argv[1]) as f:
    users = {u['id']: u for u in json.load(f)['data']}
with open(sys.argv[2]) as f:
    orders = json.load(f)['data']

rows = [{
    'order_id': o['id'],
    'customer': users.get(o['user_id'], {}).get('name', 'Unknown'),
    'amount': o['total'],
    'status': o['status'],
} for o in orders]

with open(sys.argv[-1], 'w') as f:
    json.dump({'rows': rows}, f)

Call with:

transform_data({
  language: "python3",
  script: "...",
  inputFiles: ["long_responses/users.txt", "long_responses/orders.txt"],
  outputFile: "orders-with-customers.json"
})

Filtering & Aggregation

Summarize data before display:

Python:

import json, sys
from collections import defaultdict

with open(sys.argv[1]) as f:
    data = json.load(f)

# Group by category and sum
totals = defaultdict(lambda: {'count': 0, 'total': 0})
for item in data['transactions']:
    cat = item.get('category', 'Other')
    totals[cat]['count'] += 1
    totals[cat]['total'] += item['amount']

rows = [{'category': k, 'count': v['count'], 'total': v['total']}
        for k, v in sorted(totals.items(), key=lambda x: -x[1]['total'])]

with open(sys.argv[-1], 'w') as f:
    json.dump({'rows': rows}, f)

Node.js Alternative

When Python isn't available or you prefer JavaScript:

Node:

const fs = require('fs');
const data = JSON.parse(fs.readFileSync(process.argv[2], 'utf-8'));

const rows = data.items.map(item => ({
  id: item.id,
  title: item.title,
  status: item.state,
  created: item.created_at,
}));

fs.writeFileSync(process.argv.at(-1), JSON.stringify({ rows }));

Security & Constraints

• Isolated subprocess: Scripts run in a child process with no access to API keys, credentials, or sensitive environment variables
• 30-second timeout: Scripts that exceed 30 seconds are killed
• Path sandboxing: Input files must be within the session directory. Output files must be within the session data/ directory. Path traversal attempts (e.g., ../) are blocked.
• No network access: Scripts inherit the process environment (minus secrets) but should not make network calls — use MCP tools for data fetching, then transform locally
• Blocked env vars: ANTHROPIC_API_KEY, CLAUDE_CODE_OAUTH_TOKEN, AWS_*, GITHUB_TOKEN, OPENAI_API_KEY, GOOGLE_API_KEY, STRIPE_SECRET_KEY, NPM_TOKEN

Best Practices

Decision Tree

Is the data < 20 rows?
  → YES: Inline it directly in the datatable/spreadsheet block
  → NO: Use transform_data + "src" field

Is the data already structured JSON?
  → YES: Write a simple extraction script
  → NO: Use Python's csv, json, or string parsing to structure it

Does the user need to export/download?
  → YES: Use spreadsheet block (supports .xlsx export)
  → NO: Use datatable block (better sort/filter/group UX)

Naming Conventions

• Output files: descriptive, kebab-case — stripe-transactions.json, monthly-revenue.json
• Match the context — if user asked about "Q4 sales", name it q4-sales.json

Error Handling in Scripts

• Always validate input data exists before processing
• Use try/except (Python) or try/catch (Node) for JSON parsing
• Write partial results if possible — some data is better than an error
• Keep scripts concise — complex logic is harder to debug in the 30s timeout

Script Tips

• Prefer Python for data transformation — it's the most reliable runtime for JSON/CSV processing
• Keep scripts self-contained — no pip install or external dependencies
• Use json.dump with default serialization — don't try to format numbers in the script; let column types handle rendering
• For dates, output ISO format strings (YYYY-MM-DD) — the date column type handles formatting

Troubleshooting

"Script failed (exit code 1)"

• Check the error output for syntax errors or missing imports
• Verify input files exist at the specified paths
• Make sure the script reads from sys.argv / process.argv correctly

"Output file was not created"

• Ensure the script writes to sys.argv[-1] / process.argv.at(-1) (the last argument)
• Check that json.dump / fs.writeFileSync completed successfully
• Verify the output is valid JSON

"Input file not found"

• Input paths are relative to the session directory
• Check the exact path from the tool result that produced the file
• Use long_responses/ prefix for saved tool results, attachments/ for user-uploaded files

Empty or missing rows in table

• Verify the JSON structure: must have "rows" key with an array, or be a bare array
• Check that row keys match the column "key" fields exactly (case-sensitive)
• Ensure values match expected types (numbers for currency/percent, not strings)

Table shows "Loading..." indefinitely

• The "src" path must be the absolute path returned by transform_data — do not use relative paths
• Verify the file was actually created by transform_data (check the tool result message)