pome eval - Pome

pome eval uploads a run directory you already captured and prints the score the cloud returns. It is the second half of the capture-and-evaluate split. pome run --local captures a trace without grading it; pome eval is how that trace gets a verdict. There is no local scoring — the CLI captures, the cloud grades.

Usage

pome eval [run-dir] [options]

Arguments

Argument	Required	Description
`[run-dir]`	No	An existing run directory, `runs/<task>/<run-id>`. Omit it to use the run named by `<artifacts-dir>/latest.json`.

Options

Flag	Default	Description
`--artifacts-dir <dir>`	`runs`	Directory whose `latest.json` picks the run when you pass no run directory.
`--agent <slug>`	the manifest’s `agent.slug`	Agent identity to record the evaluation under.
`--task <name>`	the run’s recorded task name	Task name to record. Falls back to `meta.json`’s `scenario` slug, then its title.
`--api-url <url>`	`https://api.pome.sh`	Control-plane base URL.

Requirements

pome eval needs a control plane serving POST /v1/eval-sessions, and credentials — pome login or POME_API_KEY.

Examples

# Capture locally, then get a verdict
pome run --local tasks/01-bug-happy-path.md
pome eval

# Evaluate a specific run
pome eval runs/01-bug-happy-path/ses_abc123

# Record it against a different agent identity
pome eval runs/01-bug-happy-path/ses_abc123 --agent triage-bot-v2

See also

pome run — --local captures the trace this command grades.
pome inspect — read a scored run in the terminal.