Skip to content
warden
Blog

AI CLI model-catalog rollout

Model-catalog reconciliation is additive and explicit. It does not run on upgrade, daemon startup, ordinary ticks, or schedules; it does not log in, update an AI CLI, send a prompt, make a paid probe, restart an agent, or change an existing session.

The operator workflow is manual and daemon-backed:

  1. Preview with warden models refresh [--aicli ID] (alias warden backend model refresh). It runs each installed AI CLI’s non-generating model discovery, submits the evidence to the daemon and prints the proposal. A preview changes no routing policy. Account identity is conveyed only as an opaque scope_fp.
  2. Inspect the operations and blocked entries, then apply with warden models refresh --apply --yes. Automation can use POST /api/v1/model-catalog/proposals, POST /api/v1/model-catalog/proposals/{proposal_id}/apply, or MCP preview_model_catalog / apply_model_catalog_proposal.
  3. Reselect a running agent with warden switch AGENT --aicli ID (catalog pick at the agent’s own role tier), --aicli ID --tier tier-N, --tier tier-N, or an explicit --aicli ID --model MODEL pin. Existing, pinned, restored and live sessions are otherwise unchanged.
  4. Inspect history with warden models history and, if needed, roll back the exact revision with warden models rollback REV --yes. Rollback is compare-and-swap policy history, not a live-session restart.

Automatic recovery from an invalid model at launch is deferred until the PTY migration and state/prompt/error detection redesign; nothing retries or substitutes a model for you.

Partial, stale, offline, malformed, unsupported, or failed evidence cannot retire a model. Seed entries are bootstrap/unknown fallback only, not proof of current availability; accepted negative evidence is not reset by seed code. Explicit model/tier/quota choices remain operator intent. On an account or profile switch, create and inspect a fresh scoped preview.

Claude, Aider, and Goose do not have dynamic entitlement discovery in this release. For a full API example, compatibility/deprecation guidance, downgrade matrix, stuck-session handling, and release checklist, see the repository’s model-catalog rollout runbook.