Copy as Markdown Ask Claude Ask ChatGPT

Data checks

A pipeline that runs at 5 am and writes the report someone opens at 8 should not be trusted blindly. A check step looks at the data before it is delivered: fixed expectations catch the obvious, an optional agent review catches the odd, and a failure stops the run before yesterday's good file is overwritten.

name: daily-sales
llm:
  default: { provider: anthropic, model: claude-opus-5-5 }

steps:
  - id: sales
    use: postgres.read
    with: { dsn: "${DATABASE_URL}", query: "select * from sales where day = current_date - 1" }

  - id: verify
    check: sales
    expect:
      rows: { min: 1000, max: 50000 }
      not_null: [order_id, total]
      unique: order_id
      accepted: { region: [north, south, east, west] }
      range: { total: { min: 0 } }
      fresh: { field: day, max_age: 2d }
      vs_last_run: { rows: 30%, sum: { total: 40% } }
    review: >
      Daily sales report for the finance team. Flag anything that looks wrong:
      totals out of the usual range, a region with no orders, odd categories.
    on_fail: stop            # or warn

  - id: report
    use: excel.write
    input: verify
    with:
      path: "smb://fileserver/finance/sales_{date}.xlsx"
      latest: "smb://fileserver/finance/sales.xlsx"
      keep: 30

notify:
  channels:
    finance: { type: teams, webhook: "${TEAMS_WEBHOOK}" }
  rules:
    - on: check.failed
      to: [finance]

The check passes the records through unchanged, so the writer reads from it (input: verify).

Fixed expectations

Deterministic and free. Each failure explains itself, with the numbers.

Expectation	Example	Fails when
rows	{ min: 1000, max: 50000 }	the row count is outside the range
not_null	[order_id, total]	a column has empty values
unique	order_id or [a, b]	a column has duplicates
accepted	{ region: [north, south] }	a column has a value outside the list
range	{ total: { min: 0, max: 10000 } }	a number is outside the range
fresh	{ field: day, max_age: 1d }	the newest date is older than max_age (30m, 6h, 1d, 2w)
vs_last_run	{ rows: 30%, sum: { total: 40% } }	the row count, or a column's sum, changed more than that since the last run that passed

vs_last_run compares with the profile of the last run whose check passed, kept in the run store. On the first run there is nothing to compare with, so it passes.

Agent review

With review:, an agent step looks at the data too. It does not read every row: it gets a profile (row count; nulls, distinct values, min, max, mean and sum or the most common values for each column; 15 sample rows), the profile of the last delivered run, and the fixed checks that failed. That keeps the cost to one small call per run.

The agent answers with a verdict and findings in plain language, for example "The north region has no orders today; on the last run it had 312." When it passes, its findings are kept as notes. Its tokens count in the run's budget and in the step's spend.

When a check fails

• on_fail: stop (default): the step fails with the findings, so nothing after it runs and the last good output stays in place.
• on_fail: warn: the data goes on and the findings are reported.

Both emit check.failed with findings (a list of {check, detail}), so any notification rule can route them; a passing check emits check.passed. Both show on the dashboard's timeline.