fix(opencode): support interleaved reasoning_content for anthropic protocol non-Claude models

[llc1123]

Issue for this PR

Closes#14638

Type of change

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

What does this PR do?

When non-Claude models (e.g. kimi-k2.5) are used via the Anthropic SDK (@ai-sdk/anthropic / @ai-sdk/google-vertex/anthropic) with interleaved reasoning enabled, previous assistant messages containing reasoning content fail to round-trip correctly. The Anthropic SDK doesn't natively support a reasoning_content field on messages — it only serializes thinking blocks.

This PR fixes it with a two-phase approach:

1. transform.ts: Converts stored reasoning parts into thinking blocks with a placeholder signature, so the Anthropic SDK will serialize them into the request body.
2. provider.ts (fetch wrapper): Intercepts the outgoing HTTP request, extracts the thinking blocks from assistant messages, and moves them to the top-level reasoning_content (or reasoning_details) field that the upstream API expects.

Note: This is a workaround. The proper fix is proposed in #14641, which adds upstream SDK support for spreading providerOptions.anthropic onto assistant messages (vercel/ai#12760). Once that lands, this workaround can be reverted in favor of the cleaner approach.

How did you verify your code works?

• Typecheck passes (bun turbo typecheck — all 12 packages green)
• Test suite passes (1135 pass, 1 unrelated timeout)
• Manually verified with kimi-k2.5 via Anthropic SDK: reasoning content from previous turns is now correctly sent as reasoning_content on assistant messages

Screenshots / recordings

N/A — no UI changes.

Checklist

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

[Copilot]

Pull request overview

This PR fixes a bug where non-Claude models (e.g., kimi-k2.5) using the Anthropic SDK with interleaved reasoning fail to correctly round-trip reasoning content from previous assistant messages. The fix implements a two-phase transformation approach: first converting internal reasoning parts to thinking blocks with a placeholder signature (so the Anthropic SDK will serialize them), then intercepting the HTTP request to move those thinking blocks to the top-level reasoning_content/reasoning_details field that the upstream API expects.

Changes:

• Added detection and special handling for Anthropic SDK models with interleaved reasoning capability
• Implemented reasoning-to-thinking-block transformation in transform.ts using a placeholder signature
• Added fetch wrapper in provider.ts to convert thinking blocks back to reasoning_content field in HTTP requests

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 6 comments.

File	Description
packages/opencode/src/provider/transform.ts	Adds logic to detect Anthropic protocol models and inject reasoning parts with placeholder signature to trigger thinking block serialization
packages/opencode/src/provider/provider.ts	Adds fetch wrapper to intercept HTTP requests and move thinking blocks to top-level reasoning_content/reasoning_details field

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[llc1123]

Addressing Copilot's review comments:

1. JSON.parse try-catch — The existing OpenAI itemId block at L1116 does the same JSON.parse(opts.body as string) without try-catch. This is SDK-generated JSON, and the project style guide says "avoid try/catch where possible."

2 & 3. Empty content after filtering — Assistant messages always contain text or tool-call parts alongside thinking/reasoning blocks. A message with only reasoning and no text/tool-call is not a real state. The early return at L146 (if (!reasoningText) return msg) already handles the no-reasoning case.

4. Reasoning order — Intentional. The Anthropic API expects thinking blocks before text content in a message.

5. Tests — Acknowledged, can be added as a follow-up.

6. Placeholder signature fragility — Already documented in the code comment at L150-151. This is a known trade-off with no better alternative short of patching the SDK.

[github-actions]

Closing this pull request because it has had no updates for more than 60 days. If you plan to continue working on it, feel free to reopen or open a new PR.