[vglafirov]

Issue for this PR

Closes #51081

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Variant.resolve only builds variants for packages listed in PROTOCOLS. gitlab-ai-provider wasn't listed, so every GitLab model resolved to [] and lost the reasoning_options declared in models.dev.

This adds a gitlab protocol that routes on the model's backend and emits settings in the shape gitlab-ai-provider reads from providerOptions.gitlab (the option key already resolves to gitlab in aisdk.ts):

• Claude (adaptive, e.g. Opus 4.8+/5.x, Sonnet 5, Fable): thinking: { type: "adaptive", effort }, which the provider serializes to thinking: { type: "adaptive", display: "summarized" } + output_config.effort. The toggle "off" state is thinking: { type: "disabled" }. Always-thinking models (Fable) get no toggle, same as anthropicMessages.
• Claude (manual, ≤ 4.5): budget variants using thinking: { type: "enabled", budgetTokens }. Unlike direct Anthropic Opus 4.5, no effort is sent, because the provider's thinking config can't combine the two.
• GPT / Codex: reasoningEffort via openaiChat. The provider maps this to reasoning_effort (Chat) or reasoning.effort (Responses) itself.
• duo-workflow-*: no variants (server-side agent).

Which GitLab models get variants is still controlled by models.dev reasoning_options. Models with an empty list stay empty.

How did you verify your code works?

• Added spells GitLab Duo variants for gitlab-ai-provider routes to packages/core/test/variant.test.ts. The variant, GitLab plugin, models-dev, provider, and aisdk tests pass, and bun run typecheck in packages/core is clean.
• Ran Variant.resolve against the live https://models.dev/api.json GitLab catalog. For example, duo-chat-opus-5-5 → low…max with {"thinking":{"type":"adaptive","effort":"max"}}, duo-chat-sonnet-5 → none,low…max, duo-chat-gpt-6-sol → none,low…max with {"reasoningEffort":"max"}, and models with reasoning_options = [] → none.

Screenshots / recordings

N/A (no UI change)

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

[vglafirov]

Rebased onto current v2. The only conflict was the @ai-sdk/cohere entry in PROTOCOLS, which went away when Cohere moved to the native provider (#52633). The variant logic is unchanged.

I re-checked it against the bundled models.dev snapshot on v2. Without this change, all 30 GitLab models resolve to no variants. With it, the 15 that declare reasoning_options get them. That now includes the new duo-chat-sonnet-5-5 and duo-chat-gpt-6-1-sol, which become selectable with the gitlab-ai-provider 6.19.0 bump in #53352. bun typecheck passes, and the variant, GitLab plugin and aisdk tests give 71 pass, 0 fail.