ADR 0026: Workflow Condition Step¶

Date: 2026-11-13

Status¶

Accepted

Related: ADR 0021 (workflows: each step type beyond v1 gets its own decision), ADR 0024 (proposes a Branch step).

Context¶

A workflow branches today with a model decision, or on the outcome of a deploy or agent step. To branch on data that a step already returned, such as the risk that an agent reported, a workflow must ask a model. That costs a call, can be wrong, and is slow.

References render in two ways (ADR 0021, References). A value that is one reference, such as {{ triage.result.count }}, keeps its type. Any other template renders as text. So a comparison gets typed values and text, and it must coerce.

Decision¶

Add a condition step type in the logic category. It compares values and takes one of two ports, True or False. It makes no call and has no side effect. Its default timeout is 10 seconds, with no retries.

Parameters¶

conditions: 1 to 10 rows. Each row is a, operator, and b. a and b hold values or templates. b is required, except for is_empty and is_not_empty.
match: all (every row holds) or any (one row holds). Default all.
ignore_case: compare text in any case. Default false.

a and b accept a string, a number, a boolean, a list, an object, or null. Save sees a template string. The engine validates the parameters again after it renders them, and then a value that is one reference has its own type.

Operators and coercion¶

Operator	Rule
`equals`, `not_equals`	Two numbers compare as numbers. A boolean compares as `true` or `false`, in any case. Other values compare as text
`greater_than`, `greater_or_equal`, `less_than`, `less_or_equal`	Both sides are numbers or numeric text. A boolean, NaN, and infinity are not numbers
`contains`, `not_contains`	A list `a` contains `b` as an item, with the `equals` rule. A text `a` contains `b` as a substring
`starts_with`, `ends_with`	`a` is text
`matches`	`a` is text. `b` is an RE2 expression
`is_empty`, `is_not_empty`	Null, empty text, an empty list, or an empty object is empty

Numbers compare as decimals, so "0.1" equals 0.1. A skipped step renders as empty text in a template and as null in a whole reference, so is_empty holds for both.

matches uses RE2 (google-re2). Its time is linear in the input, so a pattern from a trigger payload cannot hang the engine. RE2 has no backreferences or lookaround.

Errors¶

The step evaluates every row. A row that cannot be evaluated fails the step with not_a_number, not_text, or invalid_pattern, with no retry: the same values give the same error. The type has no Failed port, so the run fails. A row error is a mistake in the workflow or in the data, and a False port would hide it as a valid answer.

Type warnings¶

When a or b is one reference to an output whose type is known (the output schema of an earlier step, or an output_fields field of an agent step), and no type of it suits the operator, validation adds the warning condition_type_mismatch. The editor outlines the row in red. An example is greater_than on a string field. A text template, a trigger value, and an untyped field have no known type, so they get no warning.

One table holds the rule in each language: FITS in the registry module of the step, and its copy in the UI.

Not decided here¶

The Branch step of ADR 0024: one port per case and an Else port, over facts and work item fields. A condition step has two ports and compares references.
Expressions, such as CEL, in a condition.

Consequences¶

A workflow can branch on step output with no model call.
A type warning does not stop activation, unlike every other warning (ADR 0021, Validation). A string field that holds numbers, such as "5", compares correctly at run time, so the warning can be wrong. A value that does not fit fails the step, and the run, when it runs.
imbi-common depends on google-re2 directly. It was already installed through cel-python in the gateway.
The rule of which types suit which operator is in Python and in TypeScript. A change to one must change the other.