feat(provider): add none variant that disables thinking for DeepSeek V4

[tanishqvec]

Issue for this PR

Closes #46178

Type of change

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

What does this PR do?

reasoningVariants returns as soon as it finds an effort entry, so for any model that declares both a toggle and effort values the toggle is silently dropped. models.dev declares both for deepseek-v4-*, so --variant none was accepted but did nothing — the model still returned reasoning. That early return is also non-undefined, so the reasoningVariants(model, base) ?? variants(base) fallback at provider.ts:1310 never runs for these models either.

This keeps the toggle's none variant alongside the effort variants, and sends the switch each provider actually honours. Zen ignores thinking (even thinking: "nonsense" returns 200) but accepts reasoning_effort: "none"; api.deepseek.com is the reverse — its docs disable thinking with {"thinking": {"type": "disabled"}} and list reasoning_effort as low|high|max only.

The effort branch deliberately returns a plain object rather than going through nonEmptyVariants, because the existing "does not replace unsupported effort options with heuristic variants" test relies on {} being returned to block the fallback. Scoping follows topP() at transform.ts:555 rather than matching on model id, since around 30 other providers resell these models over openai-compatible and may accept neither switch.

How did you verify your code works?

Ran opencode run --model opencode/deepseek-v4-flash --variant none --format json --thinking -- "What is 17*23? Answer with just the number." against this branch and against unpatched dev.

Before: a {"type":"reasoning","part":{"text":"The user is asking for the product of 17 and 23. Let me calculate: 17 * 23 = 391."}} part is still emitted. After: no reasoning part is emitted. Control on the patched build with the default variant still emits one, so the variant is what changes the behaviour.

I could not exercise api.deepseek.com directly — that account has no balance — so the official-provider payload comes from https://api-docs.deepseek.com/guides/thinking_mode rather than a live response.

Tests: bun test test/provider/transform.test.ts 429 pass (3 new), bun test test/provider/provider.test.ts 102 pass, bun typecheck clean, all from packages/opencode.

Screenshots / recordings

Not a UI change.

Checklist

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

[github-actions]

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

Based on my search, I found one related PR:

Related PR:

• PR feat(provider): add none reasoning variant for DeepSeek V4 #41955 - feat(provider): add none reasoning variant for DeepSeek V4
  • This PR is explicitly mentioned in the current PR's description as a related effort that patches the same function but sends reasoning_effort: "none" for every provider
  • The current PR (feat(provider): add none variant that disables thinking for DeepSeek V4 #46181) improves upon feat(provider): add none reasoning variant for DeepSeek V4 #41955 by handling both the thinking and reasoning_effort approaches scoped by provider, rather than applying one solution globally

This is not a duplicate but rather a related/predecessor PR that addresses the same issue with a different approach. The current PR provides a more nuanced solution that works correctly for multiple providers.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍