Model plans: what runs, and on which model
A Model tiers preset (Admin > Model tiers) only picks a model for each step; it never turns a step on or off. Which steps run depends on the ticket: a ticket's Lite and Manual options skip some, and Lite also moves the evaluator to Haiku.
Economy, with Lite and Manual both on
Economy: Sonnet plans, Haiku builds, the checks stay on Sonnet. With Lite and Manual on, the close is the lightest canon allows, and it never drops below the binding evaluator.
| Step | Runs? | Model |
|---|---|---|
| Start, Plan | yes | Claude Sonnet 5.5 (Economy's Plan) |
| Spec check | skipped (Lite) | — |
| Approve the plan | yes, you | — |
| Build | yes | Claude Haiku 5.5 (the session switches after approval) |
| Your test | yes, you (Manual only) | You test by hand; sprint manual-qa records what you saw |
| Code simplifier, code reviewer, doc audit, skill eval, break-it | skipped (Lite) | — |
| Reviewer | skipped (Lite and Manual) | — |
| Security review | yes, always | Runs in your session, on its current model: Haiku after the build switch |
| Mutation test | only if logic files changed | Skipped with a reason otherwise |
| Evaluator | yes, always | Claude Haiku 5.5, not Economy's Sonnet: Lite runs the evaluator on Haiku. Manual makes it a read-only audit of your test record: it runs no tests, and checks your record is fresh, has no failing step, and matches the diff and the acceptance |
| Close | yes | — |
In short: two models (Sonnet to plan, Haiku for the rest), your own test as the main evidence, and the binding evaluator still runs, on Haiku.
What else can change a run's models
- The agent the session starts with. Admin's Agent picker only selects which agent's picks you are viewing; a session started with Copilot CLI uses the Copilot CLI column.
- A ticket's own lines.
Gate model:(Plan tab) overrides the Spec check, Reviewer and Evaluator for that ticket; aBuild model:line in its Sign-off overrides Build. - Second eval runs only when it has a model and the ticket is high-risk or says
Second eval: on. While it is Off it never runs.
The steps and their skip rules come from the project's pipeline (Factory > Pipeline, which also shows each step's model).