Show and tell: OpenAI Responses tool calls for WorkPaper spreadsheets

[gregkonush]

I added a small OpenAI Responses tool-call wrapper for Bilig WorkPaper.

The use case is narrow: an agent writes workbook inputs, reads the calculated values the workbook produced, and saves the WorkPaper document as JSON. The workbook stays application-side. The model only receives explicit tool results such as edited cell, previous value, calculated value, persisted bytes, restore checks, and verification status.

This is intentionally an explicit tool contract, not terminal control and not screen automation. It is useful when a service or agent needs formulas, cell addresses, readback, and durable state without driving Excel, Sheets, or a browser grid.

Start here:

• Docs: https://proompteng.github.io/bilig/agent-workpaper-tool-calling-recipe.html#openai-responses-api-tool-wrapper
• Example adapter: https://github.com/proompteng/bilig/blob/main/examples/headless-workpaper/openai-responses-tool-wrapper.ts
• Framework adapter examples: https://github.com/proompteng/bilig/blob/main/examples/headless-workpaper/agent-framework-adapters.ts
• Compatibility notes: https://proompteng.github.io/bilig/where-bilig-is-not-excel-compatible-yet.html

No-key package checks:

npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door workpaper-service --json
npm exec --yes --package @bilig/workpaper@latest -- bilig-evaluate --door agent-mcp --json

Feedback that would help:

• Does this shape fit how you wire OpenAI Responses tool calls today?
• Should the example show streaming function_call_output events, not just the final loop?
• What should an agent return after a write: value only, formula/value pair, changed dependents, or a small audit object?
• If you have tried spreadsheet automation from an agent, what was the first blocker: formulas, imports, persistence, cell addressing, or trust in the result?

I am not presenting this as an Excel clone. The goal is a practical workbook runtime for TypeScript services and agent tools that need formulas, edits, readback, and durable state.