Building an acceptance harness for AI spreadsheet edits: allowlists, recalc checks, idempotency

ukdicha-modakkk · reddit · 2026-10-16

The author argues "only update the forecast inputs" is an incomplete acceptance test: an agent can overwrite formulas with constants, change chart source ranges, or edit neighboring assumptions while totals still look plausible.

Define the contract before editing:

Which cells may change, and whether each change may affect values, formulas, or formatting
Which formulas, named ranges, and source references must stay unchanged
Which chart ranges and summary links must still point to intended data
Which domain-specific relationships (totals) must hold after recalculation

Univer CLI provides the inspection primitives—it distinguishes stored values, types, formulas, and display text, plus worktree comparisons and structured diffs—but that's an ingredient for a harness, not a built-in audit. Workflow: compare draft vs baseline, inspect references, explicitly trigger recalculation, wait for completion, read resulting values, and tie changes back to input evidence. A matching total cannot substitute for cell-level accounting.

Idempotency matters too: running the same update twice must not append a second adjustment or shift a range again—that invariant belongs to the workflow you build, not the tools.

Original post →

Building an acceptance harness for AI spreadsheet edits: allowlists, recalc checks, idempotency

More from coding & agent