fix(opencode): preserve reasoning_content for Kimi K2.5/K2.6 multi-turn tool calls

[j0hk3r1]

Summary

Kimi K2.5 and K2.6 (Moonshot AI) reject subsequent requests in multi-turn tool-calling conversations with:

thinking is enabled but reasoning_content is missing in assistant
tool call message at index [N]

The model returns reasoning_content alongside tool_calls in the first turn, but opencode strips it before building the next request's message history. Kimi requires reasoning_content to be preserved on every assistant message (even empty) when thinking mode is active — which is the default on Moonshot's API, Cloudflare Workers AI, Cerebras, and any other openai-compatible Kimi K2.5/K2.6 hosting.

This is the same class of bug already fixed for DeepSeek in transform.ts.

Symptom in the wild

Observed empirically with cloudflare/@cf/moonshotai/kimi-k2.6 via the openai-compatible SDK:

• Sisyphus/orchestrator agent says "I should call Oracle now" as text
• No structured tool_call is emitted in the response
• finish_reason: stop
• Session hangs until user manually re-prompts

Multiple matching reports:

• Kimi for Coding (k2p5) fails with tool calls when thinking enabled: reasoning_content is missing in assistant tool call message #10996 — Kimi K2.5/K2p5 fails with thinking-enabled tool calls (the issue this PR fixes)
• DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors #24722 — DeepSeek thinking mode tool call 400 (same root cause, already fixed for DeepSeek)
• [DeepSeek V4] reasoning_content not round-tripped — 400 error on multi-turn tool calls #24190 — DeepSeek V4 reasoning_content not round-tripped

Fix

Two changes in packages/opencode/src/provider/transform.ts:

1. Extend the reasoning-placeholder injection guard (isKimiK2RequiringReasoning) to match Kimi K2.5/K2.6 model IDs alongside DeepSeek. Matches: kimi-k2.5, kimi-k2.6, kimi-k2p5, kimi-k2p6, kimi-k2-5, kimi-k2-6. Kimi K2-Thinking is excluded — it returns reasoning_content natively and uses a different SDK path.

2. Add an explicit interleaved-style serialization block that places reasoning_content into providerOptions.openaiCompatible for Kimi K2.5/K2.6 on @ai-sdk/openai-compatible, even when the models.dev catalog has not yet declared interleaved: { field: "reasoning_content" } for these models. This avoids depending on a separate models.dev change to fix the bug.

Tests

Added 4 test cases following the existing DeepSeek test pattern:

• Kimi K2.6 injects empty reasoning placeholder on assistant tool-call without reasoning
• Kimi K2.5 preserves existing reasoning content in providerOptions
• Kimi K2-Thinking is NOT touched (returns reasoning_content natively)
• Kimi K2 base (non-K2.5/K2.6) is not affected

All 228 existing + new transform tests pass. All 340 tests under test/provider/ pass with no regressions.

Test plan

• bun test packages/opencode/test/provider/transform.test.ts — 228 pass, 0 fail
• bun test packages/opencode/test/provider/ — 340 pass, 0 fail
• Manual: route kimi-k2.6 (via Cloudflare Workers AI or Cerebras openai-compatible endpoint) through a multi-turn tool-calling session

Fixes #10996
Related: #24722, #24190

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search, PR #28542 (the current PR) does not have exact duplicates, but there are related PRs that address similar reasoning_content preservation issues:

Related PRs (Not Duplicates):

1. PR fix(provider): skip empty reasoning_content to preserve KV cache hits #28352 - fix(provider): skip empty reasoning_content to preserve KV cache hits

   • Related to the same reasoning_content handling, but focused on optimization (skipping empty reasoning to preserve cache hits)

2. PR fix(opencode): ensure DeepSeek reasoning_content round-trips for all interleaved variants #25110 - fix(opencode): ensure DeepSeek reasoning_content round-trips for all interleaved variants

   • Addresses the same class of bug but for DeepSeek models; this PR builds on that pattern

3. PR fix(provider): preserve reasoning_content on second interleaved pass (#24146 follow-up) #24443 - fix(provider): preserve reasoning_content on second interleaved pass (#24146 follow-up)

   • Another reasoning_content preservation fix, likely part of the same refactoring effort

4. PR fix(provider): auto-enable interleaved for reasoning models #24218 - fix(provider): auto-enable interleaved for reasoning models

   • Related to auto-enabling interleaved mode for models with reasoning

5. PR fix(opencode): filter empty assistant content for @ai-sdk/openai-compatible #27914 - fix(opencode): filter empty assistant content for @ai-sdk/openai-compatible

   • Potentially touches similar @ai-sdk/openai-compatible logic

These are related work but not duplicate PRs. PR #28542 is specifically targeting the Kimi K2.5/K2.6 regression, following the established pattern used for DeepSeek fixes.

[j0hk3r1]

Closing — not ready to upstream yet, will revisit after more local validation.