[Bug] Zen Go /v1/chat/completions strips usage.completion_tokens_details for deepseek-v4-flash - official DeepSeek API sends it

Description

The Zen Go OpenAI-compatible endpoint (https://opencode.ai/zen/go/v1/chat/completions) strips usage.completion_tokens_details.reasoning_tokens from responses for deepseek-v4-flash, while the official DeepSeek API (https://api.deepseek.com/chat/completions) reports it for the same model id, same prompt, same request shape.

The thinking text itself is forwarded (choices[0].message.reasoning_content is populated), so the model clearly engages reasoning through the gateway — only the token accounting for it disappears. Two additional DeepSeek-native usage fields (prompt_cache_hit_tokens, prompt_cache_miss_tokens) are likewise absent, and usage. prompt_tokens_details arrives hollowed ({} even when populated upstream).

This is a provider-specific gap in the gateway's normalization layer rather than a model-behavior difference: on this exact endpoint/harness/prompt combination, a kimi-k2.6 request returns a fully-populated details block minutes apart from a deepseek-v4-flash request that returns none.

Expected Behavior

Per the official DeepSeek API docs, the response usage should include:

"usage": {
   "completion_tokens": 0,
   "prompt_tokens": 0,
   "prompt_cache_hit_tokens": 0,
   "prompt_cache_miss_tokens": 0,
   "total_tokens": 0,
   "completion_tokens_details": {
     "reasoning_tokens": 0
}

and per the OpenAI chat-completions convention that Zen's own normalization code follows (see PR #24441), completion_tokens_details.reasoning_tokens is part of the standard
usage shape the gateway is expected to forward.

Actual Behavior

"usage": {
  "prompt_tokens": 124,
  "completion_tokens": 198,
  "total_tokens": 322,
  "prompt_tokens_details": {}
}

• usage.completion_tokens_details is absent entirely on every Zen Go response for deepseek-v4-flash (12+ attempts over two days, at different hours, trivial and reasoning-
  forcing prompts), while message.reasoning_content is delivered.
• usage.prompt_tokens_details arrives as {} (hollowed) instead of the populated shape the official API sends.
• prompt_cache_hit_tokens / prompt_cache_miss_tokens are dropped, masking whether DeepSeek prefix caching engaged per request.
• The same harness seconds apart returns a fully-populated details block for kimi-k2.6 — so the gateway is capable of forwarding it; the gap is specific to the DeepSeek path.

Plugins

No response

OpenCode version

1.18.23

Steps to reproduce

The strip is deterministic across 12+ requests spanning two days, trivial and forcing
prompts, stream=false, at different hours.

1. Send the identical request body to both endpoints (fixed combinatorial prompt that forces reasoning):

FIXED='Solve step by step internally, then respond ONLY with valid JSON {"count": <int>}. How many distinct arrangements of the letters in "MISSISSIPPI" have no two S characters   adjacent?'

BODY='{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"'"$FIXED"'"}],"max_tokens":8000,"response_format":{"type":"json_object"},"stream":false}'

# Zen Go — completion_tokens_details MISSING
curl -s https://opencode.ai/zen/go/v1/chat/completions \
  -H "Authorization: Bearer $OPENCODE_GO_API_KEY" \
  -H "Content-Type: application/json" -d "$BODY" | jq '.usage'

# Official DeepSeek — completion_tokens_details.reasoning_tokens present
curl -s https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" -d "$BODY" | jq '.usage'

2. Zen Go result (capture F1, cf-ray a31b27150f0aa330-SEA, 2026-08-27 12:41:52 GMT):

{"prompt_tokens": 124, "completion_tokens": 198, "total_tokens": 322,
 "prompt_tokens_details": {}}

choices[0].message.reasoning_content is populated (~500 chars of thinking) — the model reasoned; only the accounting vanished.

3. Official result minutes later (capture D2, 2026-08-27 11:33:14 GMT), same model id + prompt:

{"prompt_tokens": 144, "completion_tokens": 221, "total_tokens": 365,
 "prompt_tokens_details": {"cached_tokens": 0},
 "completion_tokens_details": {"reasoning_tokens": 214},
 "prompt_cache_hit_tokens": 0, "prompt_cache_miss_tokens": 144}

4. Positive control through the same gateway, same harness/prompt, minutes apart (capture F2, cf-ray a31b2b91ef3390db-SEA, 2026-08-27 12:45:43 GMT) — kimi-k2.6 returns
   the full OpenAI details shape:

"prompt_tokens_details": {"audio_tokens": 0, "cached_tokens": 47, "cache_write_tokens": 0},
"completion_tokens_details": {"audio_tokens": 0, "reasoning_tokens": 1757}

Captures were made with plain httpx/curl POSTs (no SDK, no OpenCode client parsing) — this rules out client-side field stripping

Screenshot and/or share link

No response

Operating System

macOS 14.5.0, arm64

Terminal

Ghostty

Start by running the paired curl requests against the Zen Go and official DeepSeek /v1/chat/completions endpoints, then inspect the gateway's provider-specific response normalization path. Compare the returned usage objects for deepseek-v4-flash, using the kimi-k2.6 response as a positive control; done means the documented DeepSeek usage fields are preserved in the normalized response.