[FEATURE]: Route the built-in deepseek provider through the Responses API (native support for DeepSeek-V4-Flash-0731) #39992

Problem

DeepSeek officially supports the OpenAI Responses API for deepseek-v4-flash (live since 2026-07-31; docs: https://api-docs.deepseek.com/zh-cn/guides/responses_api), but opencode's built-in deepseek provider is hardwired to Chat Completions: models.dev pins it to @ai-sdk/openai-compatible, and the OpenAIResponses protocol (packages/llm/src/protocols/openai-responses.ts) is only wired for openai/xai/azure/github-copilot/amazon-bedrock.

Verified workaround

Define a custom provider in opencode.json whose ID is a responses-routed built-in (e.g. meta) with @ai-sdk/openai + baseURL: https://api.deepseek.com/v1. Requests then hit POST /v1/responses with reasoning: {effort} and KV-cache hits (verified via wire capture). Drawbacks: wrong effort tiers, no reasoning_options in config models, unrelated catalog models.

Known gaps

1. @ai-sdk/openai 3.0.84 never parses response.reasoning_text.delta → DeepSeek thinking text is silently dropped on the AI SDK path (model still reasons; reasoning_tokens are accounted). Only the native runtime (OPENCODE_EXPERIMENTAL_NATIVE_LLM=1 + provider ID openai/anthropic/opencode*) displays it.
2. Native OpenAIResponses protocol rejects effort max (OpenAI-spec-shaped; openai-options.ts:5-8 filters it) — but DeepSeek supports max → max and maps xhigh/medium → high.
3. deepseek-v4-pro Responses support lands Aug 2026; only deepseek-v4-flash today. Responses API is stateless (store/previous_response_id ignored).

Requested

1. Responses route for the deepseek profile in packages/llm (following providers/openai.ts: responses + chat selectors).
2. sdk.responses() loader for deepseek on the AI SDK path.
3. DeepSeek's official effort set none/low/high/max (default high) — the mapping table is at https://api-docs.deepseek.com/zh-cn/guides/thinking_mode.
4. Keep /v1/chat/completions working.

Related: #35732 (custom-provider Responses path), #39933 (chat-completions reasoning toggle), #39829 (opencode-go Responses endpoint).