feat: chain Responses API response IDs

[joaomj]

Issue for this PR

Closes #34133

Type of change

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

What does this PR do?

The Responses API returns a response ID that can be used on the next provider turn as previous_response_id. OpenCode was capturing that metadata but always serializing the full conversation again.

This change carries the response ID through the prompt loop and provider request options. Once a Responses API response ID is available, only the newest message is sent on the next continuation turn. The chain is reset after compaction and whenever processing does not continue. Other providers keep their existing behavior.

How did you verify your code works?

• Ran bun test test/provider/transform.test.ts in packages/opencode: 348 tests passed.
• Ran bun x tsgo --noEmit in packages/opencode: passed.
• The repository pre-push typecheck is currently blocked by an unrelated existing packages/ui error involving missing exports from motion.

Screenshots / recordings

Not applicable, this is a backend/session runtime change.

Checklist

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

[github-actions]

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

[Tuxie]

I hope that isResponsesApiModel() doesn't prevent previousResponseId from being set when using GPT 5.6 via a personal LiteLLM or similar proxy, or for other models that benefit from previous_response_id, such as Grok 4.5 and Muse Spark 1.1

[joaomj]

I hope that isResponsesApiModel() doesn't prevent previousResponseId from being set when using GPT 5.6 via a personal LiteLLM or similar proxy, or for other models that benefit from previous_response_id, such as Grok 4.5 and Muse Spark 1.1

yes... LiteLLM works when configured with @ai-sdk/openai, but the default @ai-sdk/openai-compatible path is not currently chained. There is also an xAI allowlist mismatch: it receives previousResponseId without switching to delta-only messages. I’ll align these checks and add coverage before merge.