BrowserCortex workflows are finite JSON programs interpreted by trusted operations. They never contain JavaScript, shell commands, arbitrary functions, recursive calls, or unbounded loops.
A workflow includes a schema version, ID, semantic version, name, typed input and output descriptions, origin scope, capabilities, source dependencies, tool-version dependencies, limits, and ordered steps. A step has a stable ID, one known operation, bounded input references, optional output name, timeout, failure policy, and validation rules.
The initial operations are source.select, csv.parse, rows.filter, rows.sort, rows.project, rows.aggregate, records.compare, memory.search, text.extract, draft.create, preview.show, approval.require, tool.invoke, and file.export.
Validation performs six passes:
- Parse and validate the versioned schema.
- Resolve types and backward-only references.
- Resolve source and trusted tool dependencies.
- Derive capabilities and effect risk from trusted operation definitions.
- Enforce step, row, time, output, and nesting limits.
- Produce a canonical plan fingerprint and preview.
Currency uses safe integer minor units or canonical decimal strings. A monetary aggregate declares a currency guard with an ISO-style three-letter currency field and a currency output; mixed currencies stop the operation instead of being silently combined. Aggregate results remain numbers only when they are safe integers; fractional or larger exact results are canonical decimal strings instead of rounded binary floats. Sorts use a fixed en primitive collation by default, exact Decimal comparison with mode: "decimal", or canonical YYYY-MM-DD calendar comparison with mode: "iso-date". Date-like values are rejected unless the date mode is declared. CSV export protects formula-like cells and reports escapes. An embedded approval.require step requests approval from the broker; it never approves itself.