Test logging in Multimeter, the AI-powered REST Client for VS Code.
Markdown (no JavaScript)
# Test logging ### Checks and asserts The log level for each check/assert result follows the [report levels](../../files/test/reports.md#report-levels-all-fails-none) (`all` / `fails` / `none`): | Report level | On fail | On pass | |-------------|---------|---------| | `all` | `error` | `info` | | `fails` | `error` | `debug` | | `none` | `debug` | `trace` | Default report levels: - **Direct run** (`internal`): `all` — failures log at `error`, passes at `info` - **Imported or in suite** (`external`): `fails` — failures log at `error`, passes at `debug` This means during a direct test run you see all check results prominently, while imported test checks only surface failures unless you lower the log level. ```yaml # Example: custom report levels per check steps: - call: myAPI - check: actual: _.status expected: 200 report: internal: all # direct run: pass=info, fail=error external: none # imported: pass=trace, fail=debug ``` ### Print steps | Context | Level | |---------|-------| | Direct run | `info` | | Imported or in suite | `debug` | ### Network requests (call steps) When a test calls an API via a `call` step or a `judge` step, the request and response (headers, body) are logged at `trace` level. Auth headers (`Authorization`, `x-api-key`, …) are redacted. Lower the log level to `trace` to see full network details during test runs. Judge panel **Refresh status** (list-models) also writes the same trace lines to the Multimeter Output channel.
For AI agents · llms.txt · Home