[StudioPlatforms]

Adds AI Stupid Level (aistupidlevel), an OpenAI-compatible router at https://aistupidlevel.info/v1 ($AISTUPIDLEVEL_API_KEY).

The model ids are routing modes. Each request goes to the model that currently scores best on continuously re-run coding, reasoning and tool-use benchmarks, called with the user's own provider keys (OpenAI, Anthropic, Google, DeepSeek, Moonshot, Zhipu). For Crush:

• auto-coding is the default large model;
• auto-fastest-quality (the fastest model within 5 points of the best) is the default small model;
• auto-tooling routes on a tool-use benchmark, the one that matters for agents.

Changes:

• internal/providers/configs/aistupidlevel.json: seven routing modes.
• internal/providers/providers.go: embed and registry entry, alphabetical.
• pkg/catwalk/provider.go: InferenceProviderAIStupidLevel, added to KnownProviders.

Notes on the values:

• Costs are 0, as for Copilot. It is BYOK: the provider of the chosen model bills the user's own key, so there is no per-token price to publish.
• context_window 262,144 is the smallest across the models the router can choose, so it holds whichever model is picked.
• supports_attachments: false: the router rejects image input.
• can_reason: false: the model behind a mode changes, so there is no single reasoning control to declare.

go test ./... passes and gofmt -l is clean.

Reviewers can test at no cost: any router key plus the header X-ASL-Sandbox: true runs real routing with a mock answer.

Disclosure: I work at AI Stupid Level. Prepared with AI assistance (Claude Code) and reviewed before submitting.