Model Execution Engine
How Critique chooses models for review, Remedy, and chat — plans, roles, and credit floors.
Critique routes different model roles for different jobs. Automated PR review and Remedy share a curated runtime catalog; Critique Chat uses a broader chat catalog tuned for conversation.
Runtime roles (review & Remedy)
| Role | Purpose |
|---|---|
| Review lead | Synthesizes specialist output into verdict and summary text |
| Review specialist | Parallel lanes (security, tests, architecture, performance) |
| Remedy | Bounded fix loops inside sandboxes |
Chat does not use the review runtime role list — pick chat models in Settings → Agents.
Chat vs review catalogs
Critique Chat (free tier) is intentionally narrow — typically Ling 2.6 Flash and DeepSeek V4 Flash only. Paid review lanes such as MiniMax M3 or Qwen3.7 Plus appear in PR review and Remedy policy pickers, not as free chat models. Saved chat preferences that pointed at review-only ids normalize to an allowed chat default.
Welcome pricing (example)
Catalog models occasionally ship with time-boxed welcome floors (for example MiniMax M3 at a reduced credit floor for a fixed UTC window, then shelf pricing). The live Models and Pricing pages show countdowns and strikethrough shelf prices — docs do not duplicate every promo window.
Plans and the catalog
Models are tagged with a minimum plan (Standard, Pro, or Ultra). Your installation’s plan determines which model IDs appear in policy dropdowns and whether a requested override is allowed.
The live catalog changes as providers ship new versions. The Models page and dashboard pickers are authoritative; documentation lists concepts, not every model id string.
Credit floors
Every review-capable model has a credit floor — the minimum remaining balance Critique expects before starting work with that model. Floors scale with capability: frontier models require more headroom than fast economical models.
Floors work together with Billing & credits so one run cannot silently exhaust a month on a single PR.
Fallbacks
If a primary model is unavailable, Critique may route to a configured fallback and record that in run telemetry. You see the effective model on the review run detail page, not necessarily the policy default.
OpenRouter and BYOK
Many deployments call models through OpenRouter. Teams can optionally supply their own OpenRouter key; billing and credit gates then follow BYOK and Billing & credits.
Practical guidance for admins
- Set repository policy defaults at installation level; override only where needed.
- Match plan tier to the models you select in policy — Ultra models on a Standard plan will not run.
- Use Usage to see which models actually drove spend after the first month.
For policy fields that reference models, see Policy fields.