evolve job

Start with the job ID printed by evolve run, or find it in the list:

evolve job list

Set JOB_ID to the full ID you want to inspect. Replace <job-id> below:

JOB_ID="<job-id>"

evolve job show "$JOB_ID"
evolve job trials "$JOB_ID"

Most job commands also accept an unambiguous ID prefix of at least eight characters. Full IDs avoid lookup requests and work for records shared with you.

Inspect results

Jobs
Trials
Tasks

evolve job list --search deepswe
evolve job list --kind all --scope org
evolve job show "$JOB_ID" --json

Option for `job list`	Meaning
`--search <text>`	Match job names and dataset names.
`--kind <job\|check\|all>`	Evaluations, task checks, or both. Default `job`.
`--scope <my\|shared\|org>`	Which records to list. Default `my`; see scope.
`--uploaded`	Only jobs imported from a Harbor job directory.
`--no-uploaded`	Only jobs run on Evolve, including regrades. Includes checks when `--kind` selects them.

--uploaded and --no-uploaded cannot be combined.job show accepts one or more IDs. With --json, one ID returns an object; several return an array.

evolve job trials "$JOB_ID" \
  --status INFRASTRUCTURE_ERROR,SCORING_ERROR
evolve job trials "$JOB_ID" --dataset deepswe

Option for `job trials`	Meaning
`--status <statuses>`	Comma-separated trial statuses.
`--dataset <name>`	Select one dataset in the job.

Statuses: QUEUED, RUNNING, SCORING, SCORED, SCORING_ERROR, INFRASTRUCTURE_ERROR, BUDGET, INDETERMINATE, CANCELLED.Open an individual result with evolve trial show.

evolve job tasks "$JOB_ID"

One row per task: trial count, statuses, reward, spend, and available task-check information.

job list, job trials, and job tasks return one page. Their JSON output contains items, nextCursor, and hasMore. Pass nextCursor as --cursor to continue. They accept the shared paging and formatting options: --limit, --cursor, --columns, --quiet, --no-trunc, and --no-headers. A COMPLETED job can include failed trials. Inspect the trial statuses as well as the aggregate score.

Compare jobs

Supply two to ten job IDs. Set OTHER_JOB_ID to another job ID from job list. The comparison shows aggregate results and a per-task matrix, with disagreements first.

evolve job compare "$JOB_ID" "$OTHER_JOB_ID"

Search traces

Search across the job’s parsed trial traces. Patterns use case-insensitive POSIX regular expressions; a plain word matches a substring.

evolve job grep "$JOB_ID" 'permission denied'
evolve job grep "$JOB_ID" 'permission denied' \
  --type tool_call \
  --json

Option	Meaning
`--type <event-type>`	Match only this exact event type. Maximum `100` characters.
`-l`, `--limit <n>`	Trial match groups per page. Default `50`, maximum `200`.
`--cursor <cursor>`	Continue from the returned `nextCursor`.

The pattern must contain 1 to 512 characters. Invalid or overly expensive patterns are rejected; narrow the pattern or add --type. Each result contains a trial ID, task name, exact match count, and up to five matching events. Use evolve trial trace --grep for that trial’s full matches. With --json, follow nextCursor while hasMore is true, even if a page has no matches.

Stop work

Whole job
One dataset

evolve job cancel "$JOB_ID"

Request cancellation. The returned job may still be winding down.

evolve job stop "$JOB_ID" --dataset deepswe

Stop that dataset’s trials while keeping the rest of the job. --dataset <name> is required. You can stop only trials from jobs you created.Output separates stopped trials, already-terminal trials, and missing IDs. If a batch fails, the partial report names how many outcomes remain unreported. Run the same command again to finish the remaining work; an unreported trial may already have stopped.

For individual trials, use trial stop.

Run again

Each command returns a new linked job and leaves the source results intact. Use the new job ID with job show or job trials.

Command	Work repeated
`evolve job resume "$JOB_ID"`	Failed or stopped trials; also queued trials if the source was cancelled.
`evolve job retry "$JOB_ID"`	All trials of a settled source job.
`evolve job retry "$JOB_ID" --failed-only`	Trials in `SCORING_ERROR`, `INFRASTRUCTURE_ERROR`, `BUDGET`, or `INDETERMINATE`.
`evolve job retry "$JOB_ID" --trial "$TRIAL_ID"`	Exactly the selected settled trial. Repeat `--trial` for more.
`evolve job regrade "$JOB_ID"`	The recorded verifier, using stored inputs. No replacement verifier.

Selection options and requirements

Resume requires a finished source job. It accepts -f, --filter-error-type <type>; repeat it to select exception types. Unknown types match nothing. A scored trial is not resumed, even when its reward is zero.Retry requires a finished source job unless you select trials explicitly. It accepts either --failed-only or -t, --trial <id>, not both. Copy full trial IDs from evolve job trials "$JOB_ID"; these flags do not expand prefixes. Explicit settled trials can be retried while other trials in the source job are still running. The selection is all-or-nothing: every selected trial must be eligible.Regrade requires a finished source job. It accepts --status <s1,s2,...> and --task <name>. It uses the recorded verifier and stored inputs; edits to the original task are not applied. Ineligible trials are skipped, and the command fails if none remain. In particular, shared-verifier, multi-step, and judge-backed trials cannot be regraded through this command. See Trials.Uploaded jobs cannot be resumed, retried, or regraded. Their original execution ran outside Evolve.

evolve job start accepts the same options as evolve run.

Download results

evolve job download "$JOB_ID" -o results/

results/
└── job-<id>/
    ├── config.json
    ├── result.json
    ├── evolve.json          Evolve's record
    └── <trial folders>/

Option	Meaning
`-o`, `--output-dir <dir>`	Parent directory. Default: current directory.
`--overwrite`	Allow replacing an existing job folder.

The download unpacks a Harbor job directory and adds evolve.json records at the job root and in each matched trial folder. The command prints the saved path and file count; --json returns them as path and files. Only the creator can manage a job’s shares.

evolve job share "$JOB_ID" --link
evolve job share "$JOB_ID" --email colleague@example.com

Inspect or revoke access:

evolve job shares "$JOB_ID"
evolve job unshare "$JOB_ID" --link
evolve job unshare "$JOB_ID" --email colleague@example.com

Both share and unshare require --link, repeatable --email <address>, or both. Email sharing sends the recipient a link. An enabled link is reused until you revoke it; enabling sharing later creates a new link. See Sharing for access rules.

Read an upload’s progress

evolve upload prints an import ID before the imported job is ready. Set IMPORT_ID to that full ID, or copy one from job imports.

evolve job imports --status RUNNING
evolve job import "$IMPORT_ID" --watch

Command	Options
`job imports`	Shared list options; `--status <QUEUED\|RUNNING\|COMPLETED\|FAILED>`.
`job import <id>`	`--watch` to follow it to a job or failure. Use the full import ID.

job import --watch --json prints one final job or error document.

Delete a job

evolve job delete "$JOB_ID"

The command asks before permanent deletion. Without an interactive terminal, it refuses unless you pass -y or --yes.

Deleting a job removes its stored trials, traces, analyses, and files. Only the creator can delete it. The job must have finished; active analyses or derived regrades can prevent deletion until they finish.

Global options apply to every command.

Start work

Inspect results

Manage resources

Inspect results

Compare jobs

Search traces

Stop work

Run again

Download results

Read an upload’s progress

Delete a job

Inspect results

Compare jobs

Search traces

Stop work

Run again

Download results

Share results

Read an upload’s progress

Delete a job