feat(llm): support GPT-5.6 prompt cache new options

[AbdoKnbGit]

Issue for this PR

Closes #36318

Type of change

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

What does this PR do?

Adds GPT-5.6 prompt caching support while keeping older OpenAI models on their existing behavior.

For the GPT-5.6 family, including gpt-5.6, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, and suffix-style variants like gpt-5.6-fast, opencode now sends OpenAI request-wide prompt cache options:

promptCacheOptions: { mode: implicit, ttl: 30m }

That lowers to OpenAI wire field:

prompt_cache_options: { mode: implicit, ttl: 30m }

The existing prompt_cache_key behavior remains in place. GPT-5.5 and older models do not receive prompt_cache_options, so they keep the old cache behavior.

This also parses cache_write_tokens in OpenAI usage responses so cache writes are counted separately from cached reads and normal input tokens.

OpenAI docs: https://developers.openai.com/api/docs/guides/prompt-caching

How did you verify your code works?

Ran these locally:

• bun test from packages/llm — all tests pass.
• bun typecheck from packages/llm — passes.
• bun test test/session-runner.test.ts from packages/core — all tests pass.
• bun typecheck from packages/core — passes.
• bun test --timeout 60000 test/provider/transform.test.ts from packages/opencode — all tests pass.

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, there is one related PR that might be worth noting:

Related PR (not a duplicate):

• PR fix(provider): improve prompt caching #35982 - fix(provider): improve prompt caching
  • fix(provider): improve prompt caching #35982
  • This PR also addresses prompt caching improvements, though it appears to be a general improvement rather than GPT-5.6 specific support.

The current PR (36320) is focused on adding GPT-5.6 specific prompt caching support with promptCacheOptions and cache_write_tokens parsing, which is distinct from the earlier caching improvements.

Conclusion: No duplicate PRs found for the GPT-5.6 prompt caching feature. PR #35982 addresses caching improvements more generally but is not a duplicate.