Skip to main content
Use evolve to run evaluations in the cloud. Add --help to any command for its options.

Run an evaluation

Select tasks, agents, models, and a spend cap.

Inspect results

Read scores, compare jobs, and inspect trials.

Check and analyze

Check task quality or analyze completed traces.

Read files and logs

Browse a running sandbox or its captured files.

​
Choose a command

skill and skills are different: one manages uploaded content; the other reads the manual shipped with your CLI.

​
Global options

Put options after the command, as in evolve job list --json. evolve --version or evolve -v prints the installed version. Bare evolve prints help.

​
List options

Paged lists return one page. Pass the returned cursor to read the next.
These options apply to paged list commands, job trials, job tasks, and job imports. auth org list and secrets list use the formatting options only. The local skills list has its own simpler output.

​
Scope

job list, analysis list, check list, session list, agent list, and skill list accept --scope. Email shares apply to jobs, analyses, checks, and sessions. Analyses follow the ownership and sharing of their source job. Use full IDs when opening records from another scope. Prefix lookup usually searches your own records.

​
IDs and values

Most positional UUIDs accept an unambiguous prefix of at least eight characters. A full ID avoids the extra lookup requests.
Use full trial IDs with analyze --trial and job retry --trial. Import IDs must also be copied in full.
A trial prefix may require reading the trials of every job in scope. On large accounts, prefer the full ID from evolve job trials "$JOB_ID". Repeat a repeatable flag once per value. Do not combine short flags.
Values may begin with -. If a value is also a recognized option, use --flag=value to make it explicit.
jobs, trials, analyses, sessions, and datasets alias their singular nouns. secret aliases secrets. ls aliases list, including files ls and auth org ls.agents is reserved and is not an alias of agent. skills is not an alias of skill.

​
Output

Without --json, lists use aligned tables in a terminal and tab-separated rows when piped. -q is command-specific: on lists it selects IDs; during a watch it suppresses progress. Runtime and API errors print an error object under --json, plus a human message on stderr. Usage errors print only to stderr. A partially completed job stop can print its partial result before an error.

​
Exit codes

A completed job can contain failed trials. A check criterion marked fail is a quality finding; an errored checker is an execution failure. check show also returns 1 when a task checker failed. For API error codes, see Errors.