# @siftline/cli (/docs/packages/cli)



```bash
npm install --save-dev @siftline/cli
export TYPESAFE_API_KEY=...
npx siftline --help
```

ESM only. Node 22.14 or newer. The API key is read from the environment and from nowhere else.

## Commands [#commands]

```
siftline test  <recipe.json> <fixtures.jsonl> [--max-in-flight <n>] [--min-accuracy <ratio>] [--json] [--quiet]
siftline label <recipe.json> [records.jsonl | -] [--rules <rules.json>] [--max-in-flight <n>] [--quiet]
```

`test` runs a Recipe over a file of Fixtures and reports one accuracy per Question, the lowest
as the Recipe's own, the count that would be Unsure, and every miss. The guide shows
[the report](/docs/guide/measure-with-fixtures).

`label` runs a Recipe over Records, from a file or from stdin, and writes one Decision line per
Record to stdout in input order. With `--rules` every Decision is routed first. The guide shows
[both](/docs/guide/label-records).

`label` never sends. A routed Decision names the Action its Rule selected, and the CLI stops
there. To send, pipe the Decisions into a short script that calls `dispatch` from
`@siftline/actions`, with the URLs and secrets in its environment.
[The guide](/docs/guide/label-records#sending) shows it:

```bash
npx siftline label recipe.json records.jsonl --rules rules.json | node send.ts recipe.json
```

## Options [#options]

| Option                   | Applies to | Effect                                                         |
| ------------------------ | ---------- | -------------------------------------------------------------- |
| `--max-in-flight <n>`    | both       | Records judged at once. Default 8.                             |
| `--min-accuracy <ratio>` | `test`     | Exit `1` when the lowest Question accuracy is below the ratio. |
| `--json`                 | `test`     | Print the `TestReport` as JSON, plus a top-level `drift`.      |
| `--rules <rules.json>`   | `label`    | Route every Decision through these Rules.                      |
| `--quiet`                | both       | Silence progress. Warnings and errors still print.             |

Progress, warnings and errors go to stderr. stdout is only ever the report or the Decisions,
so both commands compose with a pipe.

## Exit codes [#exit-codes]

| Code  | Meaning                                                                                                  |
| ----- | -------------------------------------------------------------------------------------------------------- |
| `0`   | Success.                                                                                                 |
| `1`   | Work that ran and failed: a miss under `--min-accuracy`, a Record the model refused, a Judge error.      |
| `2`   | Usage and input errors: bad arguments, a file that does not parse, a Fixture or Rule the Recipe rejects. |
| `130` | Interrupted by SIGINT.                                                                                   |

## From code [#from-code]

`run(argv, deps)` is the library door and the bin is a wrapper around it. `deps` is
`{ client, stdin, stdout, stderr, env, signal }`, so a test drives the whole CLI with a
scripted client and string streams and asserts on the exit code and what was written. The
docs do exactly that to produce what the guide shows.