Operations

A plan is a list of steps. Each step is an object with an operation, a column (or null for table-level operations), params and a rationale:

{"operation": "mask", "column": "email", "params": {"method": "hash"}, "rationale": "PII:EMAIL at match_rate=1.0"}

The operation must be one of the operations below, and only the parameters listed for it are accepted. validate_plan and execute_plan check the whole plan against this vocabulary and the table's schema before any step runs. execute_plan also requires confirm=True.

Vocabulary (v1)

Operation	Does
mask	Replace values (hash / partial / constant)
drop_column	Remove a column entirely
impute	Fill nulls (mean / median / mode / constant)
standardize_case	Normalize text casing (upper / lower / title)
trim_whitespace	Strip leading/trailing whitespace
dedupe_rows	Remove duplicate rows on a key set
bucket_numeric	Bin a numeric column (bin_width or explicit bins)
parse_date	Normalize a date/time format (format given explicitly)

Every operation has a fixed parameter schema (dataveil.core.plan.OPERATIONS) checked by validate_plan/execute_plan before anything runs.

operation_vocabulary() in dataveil.core.plan returns the same vocabulary, with each parameter's allowed values, as the dictionary an agent receives from the MCP server.

No new operations without a use case

Operations are added when a real plan needs one, not speculatively. An operation never runs free-form SQL or code. To propose one, open an issue with the plan that needs it.