Skip to content

Configurable cheaper model for the permission classifier (default deepseek flash) #313

Description

@jbellis

Allow configuring a distinct, cheaper model for the auto-permission classifier instead of inheriting the session model. Default to deepseek flash (deepseek::deepseek-v4-flash) when deepseek credentials/provider are available; fall back to the session model otherwise.

Composes with the effort-floor item in the classifier-hygiene issue: a flash-class model at low effort makes Auto mode's permission gate nearly free (the gate was measured at 65% of all LLM calls in benchmark sessions before bypass). Needs: a config surface, a provider-availability probe, and cost attribution — the permission_classifier trace record (3d3d492) already carries a model field for this.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions