halo eval steps run

Replay each harvested step against a model route and score the next action (deterministic matcher, session-clustered CIs)

Usage

halo eval steps run [flags]

Flags

Flag	Shorthand	Type	Default	Description
--judge-key-env		string		env var holding the judge’s gateway credential
--judge-model		string		pinned judge model id (with –rubric)
--judge-url		string		gateway base URL for the judge (with –rubric)
--judge-wire		string	anthropic-messages	judge wire: anthropic-messages | openai-responses
--key-env		string		env var holding the credential (x-api-key, x-goog-api-key or Bearer, by wire)
--model		string		model route: a gateway model alias, or an upstream model id (required)
--parallel		int	4	concurrent model calls
--rubric		string		optional rubric YAML: a secondary judge score per step (never changes pass)
--scorecard		string		also write the scorecard JSON to this file
--seed		uint64	1	bootstrap seed (reproducible CIs)
--suite		string		directory of harvested step cases (required)
--url		string		gateway (or upstream) base URL; https, or http to loopback (required)

Global flags

Flag	Shorthand	Type	Default	Description
--output		string	text	output format: text|json

Parent

halo eval steps