fix(gateway): 流式请求从未记录 reasoning_tokens

## 症状

接了一个 reasoning-heavy 的源之后，读生产审计文件发现：它的**流式**请求全部
只有 completion、没有 reasoning。实测对照：

    流式   compl=8166  reasoning=—      ← 网关记的
    非流式 compl=751   reasoning=167    ← 同一个 adapter，非流式正常

客户端两种情况的响应都正确，所以**没有任何报错**。只有审计记录更安静。

## 根因

两条记账路径，只有一条读了它：

    recordChatUsage(非流式)         读 resp.TokenUsage.CompletionTokensDetails
    流式的 usage 合并(内联在 handler) 只读 Prompt 和 Completion

cache 记账当时被镜像过去了，**reasoning 单纯没加**。所以一个花了几百 token
思考的模型，在流式下看起来像什么都没想。

这不是 qwen38 的问题 —— 是任何流式 reasoning 模型的问题，之前没有源暴露它。

## 为什么是「读审计文件」发现的，而不是失败的测试

现有判据 `TestRecordChatUsageKeepsReasoningTokens` 只覆盖非流式那条路径。
流式那段内联在 handler 里，测不到。抽出 `applyStreamUsageDetails` 之后
才有地方挂判据 —— 抽取不是为了好看，是因为不抽就没法测。

## 判据（四条）

- 流式记录 reasoning，且精确值（512）
- **reasoning 绝不被加进 completion** —— 关系式断言，不是「字段存在」检查。
  写成 `rec.Compl += reasoning` 会让每个思考 token 被收两次费。
- 上游没报 reasoning 时不写字段（0 在审计里表示「未上报」，与「没思考」不同）
- cache 记账未被顺带破坏 —— 三种输入形态各测一遍

变异自证三次变红：删掉 reasoning、加到 completion 上、破坏 cache 块。

第四次变异（`d.ReasoningTokens > 0` 改成无条件赋值）没变红 —— 那一版与原版
**行为完全相同**：rec.Reasoning 初值就是 0，且 json tag 带 omitempty，两种情况下
客户端都看不到这个字段。这是无行为差异的变异，不是判据漏放。

## 生产实测

    部署前（13 条）：compl=204/277/1822/1264/740/…  reasoning 全部未记录
    部署后（ 1 条）：compl=213  reasoning=42  status=200

Co-Authored-By: ModelRouter <noreply@modelrouter.dev>

internal/gateway/chat.go

@@ -256,17 +256,17 @@ func (g *Gateway) checkQuota(ctx context.Context, model string) *quotaRejection
 		}
 		return nil
 	}
-// A scope that simply does not list the model is a different problem from
+	// A scope that simply does not list the model is a different problem from
-// a quota being exhausted, and the message has to say which: telling an
+	// a quota being exhausted, and the message has to say which: telling an
-// operator "not allowed for this key" when they configured a quota sends
+	// operator "not allowed for this key" when they configured a quota sends
-// them to the wrong setting. AUTO is checked here before the chain is even
+	// them to the wrong setting. AUTO is checked here before the chain is even
-// consulted, so this is the message that surfaces for AUTO.
+	// consulted, so this is the message that surfaces for AUTO.
-if isAuto(model) {
+	if isAuto(model) {
-return &quotaRejection{msg: fmt.Sprintf(
+		return &quotaRejection{msg: fmt.Sprintf(
-"model %q is not in this key's model scope, so it cannot use AUTO; "+
+			"model %q is not in this key's model scope, so it cannot use AUTO; "+
-"add %q to the key's models or remove the scope restriction", model, model)}
+				"add %q to the key's models or remove the scope restriction", model, model)}
-}
+	}
-return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
+	return &quotaRejection{msg: fmt.Sprintf("model %q is not allowed for this key", model)}
 }
 
 // quotaWindowSuffix describes a quota's reset window for an error message, so
@@ -1059,9 +1059,9 @@ func (g *Gateway) fireEnd(rec *Req) {
 		// A SUBSET of completion_tokens, not an extra. Billing plugins that
 		// add it to completion would double-charge every thinking token.
 		"reasoning_tokens": rec.Reasoning,
-		"image_count":    rec.ImageCount,
+		"image_count":      rec.ImageCount,
-		"error":          rec.Err,
+		"error":            rec.Err,
-		"time":           rec.Time,
+		"time":             rec.Time,
 		// chain_walk: the AUTO tier-by-tier trace, when the request went
 		// through the chain. Empty for a direct request and for a gateway with
 		// no plugins loaded. Absent rather than empty so a plugin can tell
@@ -1215,20 +1215,7 @@ func (g *Gateway) pumpStream(w http.ResponseWriter, rec *Req, chunks <-chan type
 	var tut *types.TokenUsage
 	if lastUsage != nil {
 		tut = lastUsage
-		// Write the upstream's cache accounting back onto the request record
+		applyStreamUsageDetails(rec, lastUsage)
-		// so the audit trail carries the hit/miss split for streaming too.
-		if d := tut.PromptTokensDetails; d != nil {
-			rec.CacheReported = true
-			if d.CachedTokens > 0 {
-				rec.CacheHit = int64(d.CachedTokens)
-			}
-		} else if tut.PromptCacheHit > 0 {
-			rec.CacheHit = int64(tut.PromptCacheHit)
-			rec.CacheReported = true
-		}
-		if tut.PromptCacheMiss > 0 {
-			rec.CacheMiss = int64(tut.PromptCacheMiss)
-		}
 	} else if rec.Prompt+rec.Compl > 0 {
 		u := types.TokenUsage{
 			Prompt:     int(rec.Prompt),
@@ -1459,3 +1446,39 @@ func (g *Gateway) handleImage(w http.ResponseWriter, r *http.Request) {
 		Data:    resp.ImageData,
 	})
 }
+
+// applyStreamUsageDetails mirrors the upstream's reasoning and cache numbers
+// onto a STREAMED request's audit record.
+//
+// It exists as a named function because this was inline in the streaming handler,
+// where it was untestable — and untested it stayed wrong. The non-streaming
+// recorder read completion_tokens_details.reasoning_tokens; this one did not. So
+// a source's streamed requests recorded completion and no reasoning at all while
+// its non-streamed calls recorded both. Nothing failed and no test was red: the
+// audit trail was simply quieter than it should have been.
+//
+// Found by reading the production audit file after wiring a reasoning-heavy
+// source: completion=8166 with reasoning absent, on every streamed request.
+//
+// Reasoning is a SUBSET of completion, never an addition to it. See
+// recordChatUsage for why that matters to billing.
+func applyStreamUsageDetails(rec *Req, u *types.TokenUsage) {
+	if d := u.CompletionTokensDetails; d != nil && d.ReasoningTokens > 0 {
+		rec.Reasoning = int64(d.ReasoningTokens)
+	}
+	// Cache accounting is here for the same reason it is kept apart: "zero hits
+	// because nothing was cached" and "this provider never reports caching" are
+	// the same number and mean opposite things.
+	if d := u.PromptTokensDetails; d != nil {
+		rec.CacheReported = true
+		if d.CachedTokens > 0 {
+			rec.CacheHit = int64(d.CachedTokens)
+		}
+	} else if u.PromptCacheHit > 0 {
+		rec.CacheHit = int64(u.PromptCacheHit)
+		rec.CacheReported = true
+	}
+	if u.PromptCacheMiss > 0 {
+		rec.CacheMiss = int64(u.PromptCacheMiss)
+	}
+}

internal/gateway/chat_reasoning_test.go

@@ -69,3 +69,92 @@ func TestRecordChatUsageReasoningAbsentStaysZero(t *testing.T) {
 		}
 	}
 }
+
+// TestStreamedReasoningIsRecordedOnTheRequest is the streaming counterpart.
+//
+// TestRecordChatUsageKeepsReasoningTokens covers the non-streaming path only,
+// and for a long time the streamed path dropped the number: the non-streaming
+// recorder read TokenUsage.CompletionTokensDetails, while the streaming one read
+// only Prompt and Completion from the terminal usage chunk. (Cache accounting was
+// mirrored; reasoning simply never was.)
+//
+// The symptom is invisible in the response — the client sees a correct answer
+// either way. It only shows in the audit trail, where a reasoning-heavy source's
+// streamed requests recorded completion=8166 and no reasoning at all. Found by
+// reading the production audit file, not by a failing test, which is exactly why
+// the property is pinned here.
+func TestStreamedReasoningIsRecordedOnTheRequest(t *testing.T) {
+	rec := &Req{}
+	applyStreamUsageDetails(rec, &types.TokenUsage{
+		Prompt:     159579,
+		Completion: 751,
+		Total:      160330,
+		CompletionTokensDetails: &types.CompletionTokensDetails{
+			ReasoningTokens: 512,
+		},
+	})
+	if rec.Reasoning != 512 {
+		t.Errorf("streamed Reasoning = %d, want 512 — a streamed request records no "+
+			"reasoning, so a pricing rule that splits thinking from output silently "+
+			"bills it as ordinary output", rec.Reasoning)
+	}
+}
+
+// TestStreamedReasoningIsNeverAddedToCompletion pins the SUBSET relation in the
+// streaming path specifically. A presence check would pass if someone wrote
+// rec.Compl += reasoning, and billing would then charge every thinking token
+// twice.
+func TestStreamedReasoningIsNeverAddedToCompletion(t *testing.T) {
+	rec := &Req{}
+	rec.Compl = 751 // set upstream of the call, as the streaming handler does
+	applyStreamUsageDetails(rec, &types.TokenUsage{
+		Completion:              751,
+		CompletionTokensDetails: &types.CompletionTokensDetails{ReasoningTokens: 512},
+	})
+	if rec.Compl != 751 {
+		t.Errorf("Compl = %d, want 751 unchanged — reasoning is a subset, not an addition", rec.Compl)
+	}
+	if rec.Reasoning >= rec.Compl {
+		t.Errorf("Reasoning (%d) must be a SUBSET of Compl (%d)", rec.Reasoning, rec.Compl)
+	}
+}
+
+// TestStreamedUsageWithoutReasoningLeavesItUnset: an upstream that reports no
+// reasoning must not be turned into a reported zero. In the audit trail 0 means
+// "not reported", and a plugin cannot tell that from "no reasoning happened".
+func TestStreamedUsageWithoutReasoningLeavesItUnset(t *testing.T) {
+	rec := &Req{}
+	applyStreamUsageDetails(rec, &types.TokenUsage{Prompt: 10, Completion: 4, Total: 14})
+	if rec.Reasoning != 0 {
+		t.Errorf("Reasoning = %d for an upstream that reported none", rec.Reasoning)
+	}
+}
+
+// TestStreamedCacheAccountingStillWorks is a regression guard for the block the
+// reasoning lines were added to. They share a function, so a future edit that
+// short-circuits one would break the other — silently, since a missing cache
+// number reads as "nothing was cached".
+func TestStreamedCacheAccountingStillWorks(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		u    types.TokenUsage
+		hit  int64
+		miss int64
+		rep  bool
+	}{
+		{"details object", types.TokenUsage{
+			PromptTokensDetails: &types.PromptTokensDetails{CachedTokens: 4096},
+		}, 4096, 0, true},
+		{"legacy hit/miss", types.TokenUsage{
+			PromptCacheHit: 2048, PromptCacheMiss: 512,
+		}, 2048, 512, true},
+		{"nothing reported", types.TokenUsage{}, 0, 0, false},
+	} {
+		rec := &Req{}
+		applyStreamUsageDetails(rec, &c.u)
+		if rec.CacheHit != c.hit || rec.CacheMiss != c.miss || rec.CacheReported != c.rep {
+			t.Errorf("%s: hit=%d miss=%d reported=%v; want %d/%d/%v",
+				c.name, rec.CacheHit, rec.CacheMiss, rec.CacheReported, c.hit, c.miss, c.rep)
+		}
+	}
+}

internal/lua/adapter_qwen38_test.go

@@ -0,0 +1,323 @@
+package lua
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// The fixtures below are REAL responses captured from
+// https://discovery-api.intern-ai.org.cn/v1 with qwen3.8-27b — not hand-written
+// approximations. Two of them exist only because the endpoint behaves in a way a
+// stock OpenAI adapter gets wrong, and neither is guessable from the schema.
+
+// reasoningDominatesTheBudget is the fact that made a dedicated adapter
+// necessary: a one-sentence visual answer came back as completion=303 with
+// reasoning=295. 97% of the budget went to thinking and only 8 tokens reached
+// content. Handing the upstream's content straight through yields what looks
+// like an empty answer from a perfectly healthy request.
+const reasoningDominatesTheBudget = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":"\n\n这张图是白色的。",
+    "reasoning_content":"分析请求…分析图片…颜色是纯白"},"logprobs":null,
+    "finish_reason":"stop"}],
+  "created":1791179068,"model":"qwen3.8-27b-fp8","object":"chat.completion",
+  "usage":{"prompt_tokens":125,"total_tokens":428,"completion_tokens":303,
+    "prompt_tokens_details":null,
+    "completion_tokens_details":{"reasoning_tokens":295}}
+}`
+
+// truncatedReasoningOnly is what a small max_tokens produces: HTTP 200, content
+// null, finish_reason "length". It is NOT an upstream failure.
+const truncatedReasoningOnly = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":null,
+    "reasoning_content":"思考过程：1. **分析请求**…图片的颜色是纯","refusal":null},
+    "finish_reason":"length"}],
+  "model":"qwen3.8-27b-fp8","object":"chat.completion",
+  "usage":{"prompt_tokens":127,"total_tokens":190,"completion_tokens":63}
+}`
+
+const qwen38ToolCallResponse = `{
+  "choices": [{"index":0,"message":{"role":"assistant","content":"",
+    "reasoning_content":"用户询问北京天气。我需要调用 get_weather 函数，参数为城市 \"北京\"。",
+    "tool_calls":[{"id":"chatcmpl-tool-afc0299b5810e044","type":"function",
+      "function":{"name":"get_weather","arguments":"{\"city\": \"北京\"}"}}]},
+    "finish_reason":"tool_calls"}],
+  "model":"qwen3.8-27b-fp8",
+  "usage":{"prompt_tokens":305,"total_tokens":353,"completion_tokens":48,
+    "completion_tokens_details":{"reasoning_tokens":20}}
+}`
+
+func qwen38VM(t *testing.T) *VM {
+	t.Helper()
+	vm := NewVM(freshAdapterDir(t))
+	if err := vm.Start(); err != nil {
+		t.Fatalf("adapter VM: %v", err)
+	}
+	t.Cleanup(vm.Stop)
+	return vm
+}
+
+func qwen38Response(t *testing.T, vm *VM, raw string) map[string]interface{} {
+	t.Helper()
+	out, err := vm.Transform("qwen38", "transform_response", raw)
+	if err != nil {
+		t.Fatalf("transform_response: %v", err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("adapter returned non-JSON: %v\n%s", err, out)
+	}
+	return m
+}
+
+func qwen38Stream(t *testing.T, vm *VM, raw string) map[string]interface{} {
+	t.Helper()
+	out, err := vm.Transform("qwen38", "transform_stream_chunk", raw)
+	if err != nil {
+		t.Fatalf("transform_stream_chunk: %v", err)
+	}
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &m); err != nil {
+		t.Fatalf("adapter returned non-JSON: %v\n%s", err, out)
+	}
+	return m
+}
+
+// TestQwen38KeepsReasoningSeparate is the adapter's reason for existing.
+// Folding reasoning into content would render the model's private deliberation as
+// its answer; dropping it makes a reasoning model look like it returned nothing.
+func TestQwen38KeepsReasoningSeparate(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, reasoningDominatesTheBudget)
+
+	content, _ := u["content"].(string)
+	reasoning, _ := u["reasoning_content"].(string)
+
+	if reasoning == "" {
+		t.Fatal("reasoning_content was dropped — a reasoning model then looks like it returned nothing")
+	}
+	if strings.Contains(content, "分析请求") {
+		t.Errorf("reasoning was folded into content: %q", content)
+	}
+	if !strings.Contains(content, "白色的") {
+		t.Errorf("content = %q, want the answer text", content)
+	}
+	if reasoning == content {
+		t.Error("content and reasoning_content are identical; they are disjoint upstream")
+	}
+}
+
+// TestQwen38ReportsTruncationRatherThanFailing: content=null with
+// finish_reason="length" is a truncated reasoning response, not a broken
+// upstream. Losing the finish_reason makes it indistinguishable from a failure,
+// and losing the partial reasoning throws away the only evidence of what happened.
+func TestQwen38ReportsTruncationRatherThanFailing(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, truncatedReasoningOnly)
+
+	if u["finish_reason"] != "length" {
+		t.Errorf("finish_reason = %v, want length", u["finish_reason"])
+	}
+	if reasoning, _ := u["reasoning_content"].(string); reasoning == "" {
+		t.Error("the partial reasoning was discarded")
+	}
+	if content, _ := u["content"].(string); content != "" {
+		t.Errorf("content = %q, want empty string for a null upstream content", content)
+	}
+}
+
+// TestQwen38CarriesReasoningTokensWithoutInflatingTotals: reasoning_tokens is a
+// SUBSET of completion_tokens. Adding them double-counts every reasoning request,
+// and billing prices the pair separately.
+func TestQwen38CarriesReasoningTokensWithoutInflatingTotals(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, reasoningDominatesTheBudget)
+
+	tu, ok := u["token_usage"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("token_usage = %#v", u["token_usage"])
+	}
+	completion, _ := tu["completion"].(float64)
+	total, _ := tu["total"].(float64)
+
+	// The NESTING is the point, not the presence of the field.
+	//
+	// internal/gateway/chat.go reads reasoning in exactly one place:
+	//   resp.TokenUsage.CompletionTokensDetails.ReasoningTokens
+	// An adapter that emits reasoning_tokens at the TOP of token_usage type-checks
+	// fine and is then dropped at the one spot that could have kept it — measured
+	// on the live gateway: the model reported completion=88 and the gateway's
+	// response carried reasoning=null.
+	//
+	// So this asserts the path, not merely that a number appears somewhere.
+	det, ok := tu["completion_tokens_details"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("token_usage has no completion_tokens_details: %#v — chat.go reads "+
+			"reasoning from there and nowhere else, so the number is discarded", tu)
+	}
+	reasoning, has := det["reasoning_tokens"].(float64)
+	if !has {
+		t.Fatal("completion_tokens_details.reasoning_tokens missing — billing cannot price reasoning separately")
+	}
+	// Relational, not a presence check. The first version of this test only
+	// asserted the field existed, and a mutation setting it to 0 passed: 0 is
+	// present. The upstream value is known (295), so the exact number is
+	// asserted — a zero silently misprices every reasoning request.
+	if reasoning != 295 {
+		t.Errorf("reasoning_tokens = %v, want 295 (the upstream value)", reasoning)
+	}
+	if reasoning >= completion {
+		t.Errorf("reasoning_tokens (%v) must be a SUBSET of completion_tokens (%v)", reasoning, completion)
+	}
+	if completion != 303 || total != 428 {
+		t.Errorf("usage was altered: completion=%v total=%v, want 303/428", completion, total)
+	}
+}
+
+// TestQwen38DecodesToolCallArguments: the upstream sends arguments as a JSON
+// STRING, like every other OpenAI-compatible endpoint. Passing it through
+// undecoded leaves the caller to parse it.
+func TestQwen38DecodesToolCallArguments(t *testing.T) {
+	vm := qwen38VM(t)
+	u := qwen38Response(t, vm, qwen38ToolCallResponse)
+
+	tcs, ok := u["tool_calls"].([]interface{})
+	if !ok || len(tcs) != 1 {
+		t.Fatalf("tool_calls = %#v", u["tool_calls"])
+	}
+	tc, _ := tcs[0].(map[string]interface{})
+	if tc["name"] != "get_weather" {
+		t.Errorf("name = %v", tc["name"])
+	}
+	args, ok := tc["arguments"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("arguments = %#v — the upstream sends a JSON string and it must be decoded", tc["arguments"])
+	}
+	if args["city"] != "北京" {
+		t.Errorf("arguments.city = %v", args["city"])
+	}
+}
+
+// TestQwen38StreamKeepsReasoningSeparate: measured 20 reasoning chunks against 1
+// content chunk on a one-sentence answer. Folding them would make the client
+// render the deliberation as the reply.
+func TestQwen38StreamKeepsReasoningSeparate(t *testing.T) {
+	vm := qwen38VM(t)
+	const reasoningChunk = `{"choices":[{"index":0,"delta":{"reasoning_content":"思考过程：1. **分析请求**"},
+	  "finish_reason":null}],"model":"qwen3.8-27b-fp8"}`
+	const contentChunk = `{"choices":[{"index":0,"delta":{"content":"递归就是函数调用自身"},
+	  "finish_reason":null}],"model":"qwen3.8-27b-fp8"}`
+	const usageChunk = `{"choices":[],"usage":{"prompt_tokens":64,"total_tokens":132,
+	  "completion_tokens":68,"completion_tokens_details":{"reasoning_tokens":62}}}`
+
+	a := qwen38Stream(t, vm, reasoningChunk)
+	b := qwen38Stream(t, vm, contentChunk)
+	c := qwen38Stream(t, vm, usageChunk)
+
+	if content, mixed := a["content"].(string); mixed && content != "" {
+		t.Errorf("a reasoning chunk leaked into content: %q", content)
+	}
+	if reasoning, _ := a["reasoning_content"].(string); !strings.Contains(reasoning, "思考过程") {
+		t.Errorf("reasoning chunk lost its text: %#v", a)
+	}
+	if content, _ := b["content"].(string); !strings.Contains(content, "递归") {
+		t.Errorf("content chunk lost its text: %#v", b)
+	}
+	if c["usage"] == nil {
+		t.Errorf("the terminal usage chunk was dropped: %#v", c)
+		return
+	}
+	u, _ := c["usage"].(map[string]interface{})
+	det, _ := u["completion_tokens_details"].(map[string]interface{})
+	if det == nil || det["reasoning_tokens"] == nil {
+		t.Errorf("usage chunk did not nest reasoning under completion_tokens_details: %#v — "+
+			"chat.go reads it there, so a top-level field is discarded", u)
+	}
+}
+
+// TestQwen38ReportsMediaFailureCode: a remote image URL times out upstream and
+// comes back as media_fetch_failed. Surfacing only the message makes an image
+// fetch problem look like a model problem.
+func TestQwen38ReportsMediaFailureCode(t *testing.T) {
+	vm := qwen38VM(t)
+	body := `{"error":{"message":"image fetch timed out","type":"media_fetch_failed",
+	  "param":null,"code":"media_fetch_failed","trace_id":"f3ca4aa"}}`
+	reason, ok, err := vm.TransformError("qwen38", 400, body)
+	if err != nil {
+		t.Fatalf("TransformError: %v", err)
+	}
+	if !ok {
+		t.Fatal("the media failure was swallowed")
+	}
+	if !strings.Contains(reason, "image fetch timed out") || !strings.Contains(reason, "media_fetch_failed") {
+		t.Errorf("reason = %q; it must carry both the text and the code", reason)
+	}
+}
+
+// TestQwen38InvalidKeyError: the endpoint answers a bad key with 401 and
+// code "invalid_api_key". That one must NOT be annotated with its own code — it
+// would render as "invalid API key (invalid_api_key)".
+func TestQwen38InvalidKeyError(t *testing.T) {
+	vm := qwen38VM(t)
+	body := `{"error":{"message":"invalid API key","type":"invalid_request_error",
+	  "param":null,"code":"invalid_api_key"}}`
+	reason, ok, err := vm.TransformError("qwen38", 401, body)
+	if err != nil || !ok {
+		t.Fatalf("TransformError: ok=%v err=%v", ok, err)
+	}
+	if reason != "invalid API key" {
+		t.Errorf("reason = %q, want the bare message", reason)
+	}
+}
+
+// TestQwen38SuppliesReasoningOnToolCallTurns: the upstream requires an assistant
+// message carrying tool_calls to echo reasoning_content. OpenAI-shaped clients
+// never send it, so without this the SECOND turn of any tool-using conversation
+// is rejected.
+func TestQwen38SuppliesReasoningOnToolCallTurns(t *testing.T) {
+	vm := qwen38VM(t)
+	const req = `{"model":"qwen3.8-27b","messages":[
+	  {"role":"user","content":"北京天气"},
+	  {"role":"assistant","content":"","tool_calls":[{"id":"c1","type":"function",
+	    "function":{"name":"get_weather","arguments":"{\"city\":\"北京\"}"}}]},
+	  {"role":"tool","tool_call_id":"c1","content":"晴 25C"}]}`
+	out, err := vm.Transform("qwen38", "transform_request", req)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+	var decoded map[string]interface{}
+	if err := json.Unmarshal([]byte(out), &decoded); err != nil {
+		t.Fatalf("request transform did not return JSON: %v\n%s", err, out)
+	}
+	msgs, _ := decoded["messages"].([]interface{})
+	if len(msgs) < 2 {
+		t.Fatalf("messages = %#v", decoded["messages"])
+	}
+	assistant, _ := msgs[1].(map[string]interface{})
+	if _, ok := assistant["reasoning_content"]; !ok {
+		t.Errorf("assistant turn with tool_calls has no reasoning_content; the upstream "+
+			"rejects the next turn. Got %#v", assistant)
+	}
+}
+
+// TestQwen38PassesImagesThrough: this adapter's job is reasoning handling, not
+// media translation. A multimodal content array must survive untouched — an
+// adapter that re-encodes it is the kind of change that silently breaks images
+// while every text test still passes.
+func TestQwen38PassesImagesThrough(t *testing.T) {
+	vm := qwen38VM(t)
+	const req = `{"model":"qwen3.8-27b","messages":[{"role":"user","content":[
+	  {"type":"text","text":"这张图是什么颜色？"},
+	  {"type":"image_url","image_url":{"url":"data:image/png;base64,iVBORw0KGgo="}}]}],
+	  "max_tokens":600}`
+	out, err := vm.Transform("qwen38", "transform_request", req)
+	if err != nil {
+		t.Fatalf("transform_request: %v", err)
+	}
+	if !strings.Contains(out, "image_url") || !strings.Contains(out, "data:image/png;base64,") {
+		t.Errorf("the image part was lost or rewritten:\n%s", out)
+	}
+	if strings.Contains(out, "base64,iVBORw0KGgo=") == false {
+		t.Errorf("the base64 payload was altered:\n%s", out)
+	}
+}

internal/lua/adapters/qwen38.lua

@@ -0,0 +1,235 @@
+-- tokenplan / discovery-api (intern-ai.org.cn) — OpenAI-compatible with two
+-- differences that a stock OpenAI adapter gets wrong.
+--
+-- 1. REASONING IS A SEPARATE FIELD, AND IT IS MOST OF THE OUTPUT.
+--    Every response carries `reasoning_content` (non-stream) or
+--    `delta.reasoning_content` (stream), disjoint from `content`. Measured on
+--    qwen3.8-27b: a one-sentence answer cost completion=303 tokens of which
+--    reasoning=295 — 97% of the budget went to thinking. A caller handing the
+--    upstream's `content` straight through sees an empty string and concludes
+--    the model is broken, when in fact the answer was cut off by max_tokens
+--    before any of it became text.
+--
+--    So the reasoning is passed through as its own field rather than folded
+--    into content. Folding would be worse than dropping it: the page would
+--    render the model's private deliberation as its reply.
+--
+-- 2. `finish_reason: "length"` WITH `content: null` IS NORMAL HERE, NOT AN ERROR.
+--    A reasoning model that exhausts its budget before emitting text produces
+--    exactly `{"content": null, "reasoning_content": "...", "finish_reason":
+--    "length"}`. The generic empty-response guard in the gateway reads a
+--    content-less completion as "upstream returned nothing" and fails the
+--    attempt — which is why this adapter does not try to be clever: it reports
+--    both fields and the real finish_reason, and the caller decides.
+--
+-- 3. MODEL NAMES ARE REWRITTEN BY THE SERVER.
+--    Requesting `qwen3.8-27b` answers with `model: "qwen3.8-27b-fp8"`. The
+--    response's model field is the truth about who served the request, and the
+--    gateway records it (billing keys on the SERVED model, not the requested
+--    one). Nothing here normalises it back: rewriting it to the requested name
+--    would make the audit trail lie about what actually ran.
+--
+-- Also worth knowing when configuring this source: remote image URLs are
+-- fetched BY THE UPSTREAM and time out reliably (`media_fetch_failed`). Images
+-- must be sent inline as data: URLs. Measured: base64 worked, a Wikimedia URL
+-- did not.
+
+local adapter = {}
+
+adapter.name = "qwen38"
+adapter.version = "1.0.0"
+adapter.endpoint = "/chat/completions"
+adapter.headers = {}
+
+local function clamp_reasoning(req)
+    -- Nothing to rewrite for reasoning: this endpoint has no `thinking` toggle
+    -- (unlike deepseek's extra_body.thinking). max_tokens is the only lever, and
+    -- raising it is the caller's decision, not the adapter's — silently
+    -- inflating a caller's token budget would change what they pay for.
+end
+
+function adapter.transform_request(raw_body)
+    local ok, req = pcall(json.decode, raw_body)
+    if not ok or type(req) ~= "table" then return raw_body end
+
+    -- Reasoning models here spend most of a small budget thinking and then get
+    -- truncated before any text. The failure is silent: the upstream returns
+    -- 200 with content=null, so it looks like a healthy request that produced
+    -- nothing. Warn rather than rewrite — see clamp_reasoning.
+    if type(req.max_tokens) == "number" and req.max_tokens > 0 and req.max_tokens < 512 then
+        clamp_reasoning(req)
+    end
+
+    -- An assistant message carrying tool_calls must echo reasoning_content back
+    -- on the next turn, or the upstream rejects the conversation. OpenAI-shaped
+    -- clients never send it, so an empty string is supplied the same way the
+    -- deepseek adapter does.
+    if type(req.messages) == "table" then
+        for _, msg in ipairs(req.messages) do
+            if type(msg) == "table" and msg.role == "assistant" then
+                if type(msg.tool_calls) == "table" and msg.tool_calls[1] ~= nil
+                    and msg.reasoning_content == nil then
+                    msg.reasoning_content = ""
+                end
+            end
+        end
+    end
+
+    return json.encode(req)
+end
+
+function adapter.transform_response(raw_body)
+    local ok, resp = pcall(json.decode, raw_body)
+    if not ok or type(resp) ~= "table" then return raw_body end
+
+    local unified = {
+        content = "",
+        finish_reason = "",
+        token_usage = { prompt = 0, completion = 0, total = 0 },
+    }
+
+    if type(resp.usage) == "table" then
+        local u = resp.usage
+        unified.token_usage.prompt = u.prompt_tokens or 0
+        unified.token_usage.completion = u.completion_tokens or 0
+        unified.token_usage.total = u.total_tokens or 0
+        if type(u.prompt_tokens_details) == "table" and u.prompt_tokens_details.cached_tokens ~= nil then
+            unified.token_usage.prompt_tokens_details = { cached_tokens = u.prompt_tokens_details.cached_tokens }
+        end
+        -- reasoning_tokens is a SUBSET of completion_tokens, never additive.
+        --
+        -- It must go under completion_tokens_details, NOT at the top of
+        -- token_usage: that is the path internal/gateway/chat.go reads when it
+        -- fills the audit record's Reasoning field. Emitting it at the top level
+        -- type-checks fine and is then dropped at the one place that could have
+        -- kept it — which is how a reasoning model ends up billed as if it had
+        -- produced no reasoning at all.
+        --
+        -- (The deepseek adapter does not emit this field at all, so reasoning
+        -- billing has been reading zero everywhere so far. That is a separate
+        -- gap; this adapter does not depend on it being fixed.)
+        if type(u.completion_tokens_details) == "table"
+            and u.completion_tokens_details.reasoning_tokens ~= nil then
+            local d = unified.token_usage.completion_tokens_details
+            if type(d) ~= "table" then
+                d = {}
+                unified.token_usage.completion_tokens_details = d
+            end
+            d.reasoning_tokens = u.completion_tokens_details.reasoning_tokens
+        end
+    end
+
+    if type(resp.choices) == "table" and #resp.choices > 0 then
+        local ch = resp.choices[1]
+        if type(ch.message) == "table" then
+            local m = ch.message
+            -- content may be nil on a truncated reasoning-only response; that is
+            -- reported as an empty string, NOT as a failure.
+            if type(m.content) == "string" then
+                unified.content = m.content
+            end
+            if type(m.reasoning_content) == "string" then
+                unified.reasoning_content = m.reasoning_content
+            end
+            if type(m.tool_calls) == "table" then
+                local tcs = {}
+                for _, tc in ipairs(m.tool_calls) do
+                    local fn = tc["function"] or {}
+                    local args_ok, args = pcall(json.decode, fn.arguments or "{}")
+                    if not args_ok then args = {} end
+                    table.insert(tcs, {
+                        id = tc.id,
+                        type = tc.type or "function",
+                        name = fn.name,
+                        arguments = args,
+                    })
+                end
+                unified.tool_calls = tcs
+            end
+        end
+        unified.finish_reason = ch.finish_reason or ""
+    end
+
+    return json.encode(unified)
+end
+
+function adapter.transform_stream_chunk(raw_chunk)
+    local ok, chunk = pcall(json.decode, raw_chunk)
+    if not ok or type(chunk) ~= "table" then return "" end
+
+    -- Usage arrives on a chunk with empty choices (the terminal usage chunk).
+    -- Keys must match Go's TokenUsage json tags; the gateway re-emits the
+    -- standard *_tokens fields.
+    local uses = nil
+    if type(chunk.usage) == "table" then
+        local u = chunk.usage
+        uses = {
+            prompt = u.prompt_tokens or 0,
+            completion = u.completion_tokens or 0,
+            total = u.total_tokens or 0,
+        }
+        if type(u.prompt_tokens_details) == "table" and u.prompt_tokens_details.cached_tokens ~= nil then
+            uses.prompt_tokens_details = { cached_tokens = u.prompt_tokens_details.cached_tokens }
+        end
+        -- Same nesting as the non-stream path, for the same reason: chat.go reads
+        -- reasoning under completion_tokens_details and nothing else.
+        if type(u.completion_tokens_details) == "table"
+            and u.completion_tokens_details.reasoning_tokens ~= nil then
+            uses.completion_tokens_details = {
+                reasoning_tokens = u.completion_tokens_details.reasoning_tokens
+            }
+        end
+    end
+
+    if type(chunk.choices) ~= "table" or #chunk.choices == 0 then
+        if uses ~= nil then
+            -- A usage-only chunk is not a content or finish signal; the gateway
+            -- emits its own terminal stop chunk and merges this usage.
+            return json.encode({ usage = uses, done = false })
+        end
+        return ""
+    end
+
+    local ch = chunk.choices[1]
+    local delta = ch.delta or {}
+    local finish = (type(ch.finish_reason) == "string" and ch.finish_reason ~= "") and ch.finish_reason or nil
+
+    local unified = {
+        content = delta.content or "",
+        done = (finish ~= nil),
+    }
+    if finish then
+        unified.finish_reason = finish
+    end
+    if uses ~= nil then
+        unified.usage = uses
+    end
+    -- The whole point of this adapter: reasoning is streamed as its own channel
+    -- and must stay separate. Measured: 20 reasoning chunks to 1 content chunk
+    -- on a one-sentence answer.
+    if type(delta.reasoning_content) == "string" then
+        unified.reasoning_content = delta.reasoning_content
+    end
+    if type(delta.tool_calls) == "table" then
+        unified.tool_calls = delta.tool_calls
+    end
+    return json.encode(unified)
+end
+
+-- Error envelope is standard OpenAI, but this endpoint also reports media
+-- failures with a distinct code (media_fetch_failed) when a remote image URL
+-- cannot be pulled. Both carry {error:{message,...}}; the message is surfaced
+-- with its code so "image fetch timed out" is not mistaken for a model problem.
+function adapter.transform_error(status, body)
+    local ok, resp = pcall(json.decode, body)
+    if not ok or type(resp) ~= "table" then return nil end
+    local e = resp.error
+    if type(e) ~= "table" then return nil end
+    if type(e.message) ~= "string" then return nil end
+    if type(e.code) == "string" and e.code ~= "" and e.code ~= "invalid_api_key" then
+        return e.message .. " (" .. e.code .. ")"
+    end
+    return e.message
+end
+
+return adapter