DeepSeek provider: prompt_cache_miss_tokens not mapped to cache.write

[22xuan]

Problem

The DeepSeek API returns two cache fields in every response:

• prompt_cache_hit_tokens — tokens served from disk KV cache
• prompt_cache_miss_tokens — tokens that missed the cache

opencode correctly maps prompt_cache_hit_tokens → cache.read, but prompt_cache_miss_tokens is never recorded, resulting in cache.write: 0 for every single message.

Steps to reproduce

1. Use deepseek provider (any model)
2. Make multiple requests in the same session (multi-turn conversation)
3. Check stats: opencode stats --models
4. Check per-message data: opencode db "SELECT json_extract(data, '$.tokens.cache.read'), json_extract(data, '$.tokens.cache.write') FROM message WHERE json_extract(data, '$.role')='assistant'" --format tsv

Expected

cache.write should equal prompt_cache_miss_tokens from the DeepSeek API response, enabling calculation of cache hit rate as hit / (hit + miss).

Actual

cache.write is always 0 for every message. Cache read (hit) works fine.

Impact

Cannot calculate cache hit rate. The opencode stats display shows Cache Write: 0 which is misleading.

Environment

• opencode v1.17.15
• deepseek/deepseek-v4-pro, deepseek/deepseek-v4-flash
• Linux (WSL2 Ubuntu 22.04)

[github-actions]

This issue might be a duplicate of or related to an existing issue. Please check:

• OpenAI-compatible providers report zero tokens_cache_read despite upstream returning cached_tokens #34022: OpenAI-compatible providers report zero tokens_cache_read — similarly touches cache token accounting for DeepSeek/compatible providers, though that issue focuses on cache read tokens rather than cache.write / prompt_cache_miss_tokens.