Skip to main content
Check task quality against a rubric, server-side. <path> is one task directory or a directory of task directories, uploaded as one archive (upload_too_large past the published ceiling); -d names a published dataset instead, one or the other. Without --watch the command prints the accepted check with the hint Follow it with: evolve check show <id>; --json prints the accepted check body. Exit 0 on an accepted check; with --watch, and on check show, exit 1 when any task errored.

Options

name[@version]
Check a published dataset’s tasks instead of a local path. The checker reads the version’s retained task package, and the job page’s CHECK tab lists these checks by task.
name
default:"openrouter/deepseek/deepseek-v4.1-flash"
Model the checker runs. fireworks/deepseek-v4.1-flash is the same model on its Fireworks route. Must be on the claude roster.
value
Reasoning effort the checker runs at. Defaults to the per-model default.
path
Rubric file: TOML, YAML or JSON in the {criteria: [{name, description, guidance}]} shape. Defaults to the platform’s check rubric, eleven criteria.
path
Prompt file for the evaluator agent. Its text replaces the built-in check prompt, with {task_path}, {file_tree} and {criteria_guidance} rendered. Defaults to the built-in prompt.
provider
Sandbox provider the checker runs on. Defaults to the platform’s analysis default.
n
Max concurrent task checks, beneath the organization’s ceiling. Defaults to the ceiling.
glob
Only check tasks matching the glob. Repeatable.
glob
Skip tasks matching the glob. Repeatable.
n
Max tasks to check, after the globs, in sorted directory order.
Poll until every task check settles, then print the report. One task prints its label line, a CHECK | OUTCOME | EXPLANATION table with the evidence beneath each verdict, and the cost; several tasks print a TASK | LABEL | PASS | FAIL | N/A | UNKNOWN | COST ($) | TASK CHECK ID table with each failed task’s reason below it. With --json, NDJSON envelopes: check.accepted, check.progress per change, check.final.
With --watch: suppress the progress lines and print the final report only.

Global options

Machine-readable JSON output.
key
API key. Defaults to $EVOLVE_API_KEY.
url
API base URL. Defaults to the Evolve dashboard API.

Reading checks back

The same word, followed by list, show, trace or download, reads checks back. A task directory literally named like one of these verbs is written ./list.

check list

List your task quality checks, newest first.
n
Page size.
c
Resume from a page cursor.
keys | all | help
Choose and order columns, comma-separated. help lists the keys.
Print only ids, one per line, for piping.
Full cell content instead of one-line truncation.
Omit the header row in piped (TSV) output.
my | shared
Visibility scope: my, what you created (the default), or shared, your organizations’ rows that teammates created.
s1,s2,...
Filter by check status: queued, running, completed.

check show

Show one task quality check in full, with one result per task.

check trace

Print the checker’s own parsed transcript for one task check.
n
Skip the first N events. To resume, pass the count you already hold.

check download

Save a check as a check folder (check_report.json plus one folder per task) by its check id, one task check’s folder by its task check id, or stream one task check’s artifact.
dir
default:"checks/"
Directory to save under. A check id lands in <dir>/check-<id>/, a task check id in <dir>/check-<task>__<7 chars>/, the run’s own folder name.
Replace an existing folder of the same name.
artifact
Print one artifact of a task check to stdout instead of saving: task-check (the result document), trace-parsed, trace-stdout, trace-stderr or agent-home.
n
With --stream trace-parsed: skip the first N events.