Skip to content

Expressions

${{ … }} holds an expression. Expressions use a small, bounded profile of CEL named runbook-cel/1: enough to read the run’s data and make decisions, and nothing that loops forever or reaches outside the run. Every expression is type-checked when you validate or publish, so a wrong name or type is a diagnostic, not a surprise mid-run.

A value that is only an expression keeps the expression’s type. A value with text around an expression is a string, with the result interpolated:

if: "${{ steps.check.verdict == 'pass' }}" # a boolean
instructions: Roll out ${{ inputs.version }}. # a string
Name What it is Example
inputs.<name> The run’s inputs inputs.service
vars.<name> The run’s variables vars.result
fields.<name> The run’s fields fields.severity
steps.<id>.outputs.<name> A step’s reported outputs steps.measure.outputs.error_rate
steps.<id>.verdict A check’s verdict: pass, warn or fail steps.check.verdict == 'fail'
steps.<id>.status The step’s status (succeeded, failed, skipped, …) steps.migrate.status == 'failed'
steps.<id>.attempts How many attempts it took steps.flaky.attempts > 1
steps.<id>.actor, started_at, finished_at Who did it and when steps.review.actor
steps.<id>.outputs.decision, .comment, .decided_by An approval’s result steps.go.outputs.decision == 'ship'
steps.<id>.items A foreach’s iterations, each with its steps steps.regions.items[0].deploy.status
steps.<id>.iterations, .satisfied, .exhausted A loop’s iterations and how it ended steps.stabilize.exhausted
item (or the as name), index Inside a foreach: the current item and its position region, index
iteration Inside a loop: the iteration, from 0 iteration > 0
payload.<name> In a webhook trigger: the declared payload fields payload.status

A step’s record is available once the step has run; a value that may be missing (a skipped step’s outputs, an optional output) is null.

Compare ==, !=, <, <=, >, >= on numbers, strings, booleans and timestamps
Combine &&, ||, !
Choose cond ? a : b
Index list[0], object.property
Negate -x for numbers
Time now(), and timestamp('2026-12-01T09:00:00Z') for a literal time
Search list.exists(x, x.status == 'failed'), once per expression

That is the whole profile. There is no arithmetic, no string functions, and no list or map literals (expression.unsupported). When you need a computed value, have a step report it as an output and read the output.

  • Conditions: if, switch.cases[].when, loop.until and loop.while, and a webhook trigger’s when. These must be booleans.
  • Values: set.vars and set.fields, spec.vars starting values, spec.outputs, call.with, with, foreach.items (a list) and foreach.key, a webhook trigger’s inputs and dedupe.
  • Text: instructions, titles, approval.prompt, end.message, spec.execution.title and tags, spec.report, assignee.
rollout-gate.yaml
metadata: { id: rollout-gate }
spec:
inputs:
service: { type: string, required: true }
max_error_rate: { type: number, default: 0.01 }
vars:
outcome: pending
steps:
- id: measure
actor: agent
check: true
instructions: Measure ${{ inputs.service }}'s 5xx rate over 15 minutes.
outputs:
error_rate: { type: number, required: true }
- id: decide
switch:
cases:
- when: "${{ steps.measure.outputs.error_rate > inputs.max_error_rate }}"
steps:
- id: hold
set: { vars: { outcome: held } }
default:
steps:
- id: proceed
set: { vars: { outcome: proceeded } }
outputs:
outcome: "${{ vars.outcome }}"
error_rate: "${{ steps.measure.outputs.error_rate }}"
report: |
${{ inputs.service }}: ${{ vars.outcome }}
(error rate ${{ steps.measure.outputs.error_rate }})

An expression’s source is at most 4,096 bytes, 256 nodes and 32 levels deep, and its evaluation has a fixed cost budget. A runbook with more than 256 steps with IDs, or foreach and loop nested more than four deep, is too large to type (expression.schema_limit).