# PromQL validation

Nines uses the Prometheus parser and type checker pinned in `go.mod`.

- Input validation is shared by the CLI, API, operator and tools that call the service.
- It runs offline without contacting Prometheus or Mimir.

```sh
nines validate --offline -f examples/api-availability.slo.yaml --agent
nines rules generate --offline -f examples/api-availability.slo.yaml --tenant commerce
```

- Both `goodQuery` and `totalQuery` must be valid PromQL expressions returning instant vectors.
  - For example, `sum(increase(http_requests_total[1m]))` and `vector(100)` are valid types.
  - An incomplete `sum(`, unknown function, scalar `100`, string literal or range vector `http_requests_total[1m]` is rejected with the affected field in the diagnostic.
  - The vector requirement matches the collector and rule compiler's interval-count contract.
  - A valid type alone does not make an expression an appropriate event count.
- Provider-window selectors retain their restriction to metric names with exact label matchers and also pass through PromQL parsing.
  - This catches invalid escapes that a selector-shape check alone would miss.
  - Manual-ingestion SLOs may still omit both queries.
- Every generated recording and burn-rate alert expression is parsed and checked before the rule group is returned for export or deployment, including provider-window and operator-owned groups.
  - Interval queries are printed from their parsed syntax trees before nesting, so trailing comments cannot swallow generated syntax.
  - Canonical formatting may change a rule plan's digest; fetch and review a fresh plan before applying it.
- Offline validation does not establish that metrics exist, selectors match traffic, each input returns one aggregated series, values represent interval counts, good counts do not exceed total counts, or a query is affordable.
  - It also cannot prove compatibility with every backend version or extensions such as MetricsQL and LogQL.
  - Collection and integration diagnostics provide separate runtime evidence.
  - There is no additional `promtool` installation requirement for these checks.
- Previously accepted malformed or incorrectly typed queries now fail validation.
  - Correct them before applying a definition or starting a server that loads them from storage.
  - A successful validation receipt describes schema, PromQL types and generated-expression checks, rather than claiming telemetry or SLO correctness.