Bump bundled @ai-sdk/openai-compatible 2.0.41 -> 2.0.75 (fixes fragmented reasoning streams)

Locate the package declaration and lockfile entry that pin @ai-sdk/openai-compatible at 2.0.41. Update them to 2.0.75, then replay the captured SSE stream under Node or exercise an affected reasoning model. Done means reasoning output produces one reasoning-start and one reasoning-end event with the deltas kept contiguous.

Environment

• opencode: v1.x (desktop, bundled SDK)
• bundled @ai-sdk/openai-compatible: 2.0.41
• Model: any OpenAI-compatible reasoning model that streams reasoning_content (DeepSeek, Kimi, Qwen, ...)
• Platform: Windows

Problem

Reasoning output is fragmented in the UI: every single reasoning token opens and closes its own reasoning-start / reasoning-end cycle instead of accumulating into one contiguous reasoning block. The thinking text renders word-by-word on separate lines and flickers, instead of one continuous block.

Reproduction

1. Configure any model routed through @ai-sdk/openai-compatible that returns reasoning_content.
2. Send any prompt.
3. Observe the thinking block: each token is toggled on/off separately rather than appended to one block.

Evidence

Replaying the same captured SSE stream (model streams reasoning_content token by token) under Node, counting the emitted events:

With 2.0.41 every token becomes its own reasoning cycle; with 2.0.75 the stream is correctly kept as a single contiguous block (268 deltas appended to one part).

Root cause

Already fixed upstream in 2.0.75:

• vercel/ai commit f3c1980: fix(openai-compatible): keep reasoning streams contiguous when deltas include empty tool call arrays
• Released in: @ai-sdk/openai-compatible@2.0.75 (2026-09-09)

opencode currently bundles 2.0.41 (see #17902, where a maintainer confirmed "we use @ai-sdk/openai-compatible: 2.0.41"; the lockfile pins 2.0.41 as well), which does not include this fix.

Request

Please bump the bundled @ai-sdk/openai-compatible from 2.0.41 to the latest 2.0.75:

• This stays within the 2.x line (specificationVersion: v3) and remains compatible with the current AI SDK 6 core, so no core upgrade is required.
• It is a pure patch bump with no breaking changes.

References

• #17902 (source of the current 2.0.41 pin)
• #27987 / #22241 (related discussion about merging reasoning cycles on the opencode side)