Switching from plan to build mode with DeepSeek V4 Pro/Flash causes context loss and cache misses #35708

Description

When using DeepSeek V4 Pro or V4 Flash models, switching from plan to build mode causes:

1.Loss of conversation context from previous turns
2.Only ~2048 tokens hit DeepSeek's automatic prefix cache after the switch, the rest (potentially hundreds of thousands of tokens) are cache misses, resulting in significantly increased latency and cost

Hope that switching between plan/build modes of the same model can properly preserve context and cache.

Plugins

No response

OpenCode version

v1.17.14 Desktop

Steps to reproduce

1.Configure DeepSeek V4 Pro/Flash
2.Have several turns of conversation in plan mode
3.Switch to build mode (/build)
4.Observe: the next request only has ~2048 tokens of cached prefix, everything else misses, and partial context is lost

Screenshot and/or share link

Operating System

Windows 11

Terminal

Desktop