fix: GPT-5.6 Bedrock reasoning variants fail with SDK 4.0.158

[mmurilo]

Description

GPT-5.6 OpenAI models on Amazon Bedrock fail when a reasoning-effort variant is selected. OpenCode currently pins @ai-sdk/amazon-bedrock@4.0.158, which emits an invalid reasoning field for prefixed profiles such as global.openai.gpt-5.6-luna; Bedrock returns HTTP 400, for example Unknown parameter: reasoningConfig.

The upstream fix is published in @ai-sdk/amazon-bedrock@4.0.165. Update the dependency in both packages/core/package.json and packages/opencode/package.json, regenerate bun.lock, and verify the request.

Upstream history: AI SDK issue #19403, main fix #19420, v6 backport #19624, release merge #19752, and published v6 release #19775. The published 4.0.165 source sends nested reasoning.effort for GPT-5.x and retains flat reasoning_effort for gpt-oss.

Plugins

None

OpenCode version

v1.18.23

Steps to reproduce

1. Configure Amazon Bedrock with valid credentials.
2. Select global.openai.gpt-5.6-luna.
3. Select a non-default reasoning-effort variant.
4. Send a prompt.
5. Observe the Bedrock Converse HTTP 400 response.

Expected: the request uses additionalModelRequestFields.reasoning.effort and succeeds.

Screenshot and/or share link

Not applicable; the failure is in the provider request payload.

Operating System

macOS

Terminal

Not applicable

[mmurilo]

This follows the required Bug Report template. The requested fix is a dependency update to published @ai-sdk/amazon-bedrock@4.0.165; no OpenCode runtime provider change is expected.

[pengzh1]

Taking this: bump @ai-sdk/amazon-bedrock 4.0.158 → 4.0.165 in packages/core and packages/opencode, regenerate the lockfile, and check the reasoning variant request path against the upstream fix.