Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.
Go agentagenticaichatbotdsh-pluginembeddingsevaluationgenerative-aigolangknowledge-basellmmulti-tenantollamaopenaiquestion-answeringragrerankingsemantic-searchvector-searchwiki
fix(chat): send one completion-token field per provider (#3082)
lyingbug1c448bf72c6b
diff --git a/internal/agent/const.go b/internal/agent/const.goindex af068d645..d9de825a1 100644--- a/internal/agent/const.go+++ b/internal/agent/const.go@@ -100,9 +100,10 @@ func (e *AgentEngine) getLLMStallTimeout() time.Duration {// exactly fits by our arithmetic is the one that gets rejected.const contextSafetyTokens = 4096-// getCompletionTokenBudget is the max_tokens / max_completion_tokens sent on-// each ReAct LLM round. Unset without a sandbox is 4096; unset with a-// sandbox (write_sandbox_file / edit_sandbox_file) is 24576.+// getCompletionTokenBudget is the single completion budget for each ReAct LLM+// round. The chat layer maps it to max_tokens or max_completion_tokens per+// provider. Unset without a sandbox is 4096; unset with a sandbox+// (write_sandbox_file / edit_sandbox_file) is 24576.func (e *AgentEngine) getCompletionTokenBudget() int {configured := 0sandboxID := ""diff --git a/internal/agent/engine_test.go b/internal/agent/engine_test.goindex 90728f52c..41789a57c 100644--- a/internal/agent/engine_test.go+++ b/internal/agent/engine_test.go@@ -665,7 +665,7 @@ func TestStreamThinkingToEventBus_SetsCompletionTokenBudget(t *testing.T) {[]chat.Message{{Role: "user", Content: "test"}}, nil, 0, "sess-1")require.NoError(t, err)require.Len(t, mock.opts, 1)- assert.Equal(t, 4096, mock.opts[0].MaxTokens)+ assert.Zero(t, mock.opts[0].MaxTokens)assert.Equal(t, 4096, mock.opts[0].MaxCompletionTokens)})@@ -680,7 +680,7 @@ func TestStreamThinkingToEventBus_SetsCompletionTokenBudget(t *testing.T) {[]chat.Message{{Role: "user", Content: "test"}}, nil, 0, "sess-1")require.NoError(t, err)require.Len(t, mock.opts, 1)- assert.Equal(t, types.DefaultSmartReasoningMaxCompletionTokens, mock.opts[0].MaxTokens)+ assert.Zero(t, mock.opts[0].MaxTokens)assert.Equal(t, types.DefaultSmartReasoningMaxCompletionTokens, mock.opts[0].MaxCompletionTokens)})@@ -697,7 +697,7 @@ func TestStreamThinkingToEventBus_SetsCompletionTokenBudget(t *testing.T) {[]chat.Message{{Role: "user", Content: "test"}}, nil, 0, "sess-1")require.NoError(t, err)require.Len(t, mock.opts, 1)- assert.Equal(t, types.DefaultAgentMaxCompletionTokens, mock.opts[0].MaxTokens)+ assert.Zero(t, mock.opts[0].MaxTokens)assert.Equal(t, types.DefaultAgentMaxCompletionTokens, mock.opts[0].MaxCompletionTokens)})@@ -712,7 +712,7 @@ func TestStreamThinkingToEventBus_SetsCompletionTokenBudget(t *testing.T) {[]chat.Message{{Role: "user", Content: "test"}}, nil, 0, "sess-1")require.NoError(t, err)require.Len(t, mock.opts, 1)- assert.Equal(t, 64000, mock.opts[0].MaxTokens)+ assert.Zero(t, mock.opts[0].MaxTokens)assert.Equal(t, 64000, mock.opts[0].MaxCompletionTokens)})}diff --git a/internal/agent/finalize.go b/internal/agent/finalize.goindex fcde636d6..cab3aa51b 100644--- a/internal/agent/finalize.go+++ b/internal/agent/finalize.go@@ -103,7 +103,6 @@ Now generate the final answer:`, query, imageRequirement)messages,&chat.ChatOptions{Temperature: e.config.Temperature,- MaxTokens: budget,MaxCompletionTokens: budget,PromptCacheKey: sessionID,}, // Thinking disabled for final answer synthesisdiff --git a/internal/agent/think.go b/internal/agent/think.goindex ff2a0722c..6a83f8cf7 100644--- a/internal/agent/think.go+++ b/internal/agent/think.go@@ -252,7 +252,6 @@ func (e *AgentEngine) streamThinkingToEventBus(parallelToolCalls := trueopts := &chat.ChatOptions{Temperature: e.config.Temperature,- MaxTokens: budget,MaxCompletionTokens: budget,Tools: tools,Thinking: e.config.Thinking,diff --git a/internal/models/chat/anthropic.go b/internal/models/chat/anthropic.goindex 1690f357e..e9f7fbd6b 100644--- a/internal/models/chat/anthropic.go+++ b/internal/models/chat/anthropic.go@@ -268,10 +268,8 @@ func (c *AnthropicChat) buildRequest(_ context.Context, messages []Message, optsMessages: make([]anthropicMessage, 0, len(messages)),}if opts != nil {- if opts.MaxTokens > 0 {- req.MaxTokens = opts.MaxTokens- } else if opts.MaxCompletionTokens > 0 {- req.MaxTokens = opts.MaxCompletionTokens+ if budget := opts.CompletionBudget(); budget > 0 {+ req.MaxTokens = budget}if opts.Temperature > 0 {temperature := opts.Temperaturediff --git a/internal/models/chat/chat.go b/internal/models/chat/chat.goindex 48c85371a..5a908ee22 100644--- a/internal/models/chat/chat.go+++ b/internal/models/chat/chat.go@@ -26,11 +26,15 @@ type FunctionDef struct {// ChatOptions 聊天选项type ChatOptions struct {- Temperature float64 `json:"temperature"` // 温度参数- TopP float64 `json:"top_p"` // Top P 参数- Seed int `json:"seed"` // 随机种子- MaxTokens int `json:"max_tokens"` // 最大 token 数- MaxCompletionTokens int `json:"max_completion_tokens"` // 最大完成 token 数+ Temperature float64 `json:"temperature"` // 温度参数+ TopP float64 `json:"top_p"` // Top P 参数+ Seed int `json:"seed"` // 随机种子+ // MaxTokens and MaxCompletionTokens are aliases for one completion budget.+ // Callers may set either; CompletionBudget() prefers MaxCompletionTokens.+ // The outbound Chat Completions JSON carries exactly one of max_tokens or+ // max_completion_tokens, chosen per provider (see wireCompletionTokenField).+ MaxTokens int `json:"max_tokens"`+ MaxCompletionTokens int `json:"max_completion_tokens"`FrequencyPenalty float64 `json:"frequency_penalty"` // 频率惩罚PresencePenalty float64 `json:"presence_penalty"` // 存在惩罚Thinking *bool `json:"thinking"` // 是否启用思考diff --git a/internal/models/chat/completion_budget.go b/internal/models/chat/completion_budget.gonew file mode 100644index 000000000..fe915a75f--- /dev/null+++ b/internal/models/chat/completion_budget.go@@ -0,0 +1,72 @@+package chat++import (+ "github.com/Tencent/WeKnora/internal/models/provider"+ "github.com/sashabaranov/go-openai"+)++// completionTokenField is the single Chat Completions wire name for a+// completion budget. OpenAI treats max_tokens and max_completion_tokens as+// mutually exclusive; gateways such as Volcengine Ark reject requests that+// carry both (Tencent/WeKnora#3014).+type completionTokenField string++const (+ completionTokenFieldMaxTokens completionTokenField = "max_tokens"+ completionTokenFieldMaxCompletionTokens completionTokenField = "max_completion_tokens"+)++// CompletionBudget is the single per-call generation cap. MaxTokens and+// MaxCompletionTokens on ChatOptions are input aliases for the same budget+// (YAML, older callers, and the agent UI all feed one of them). When both are+// set, the newer MaxCompletionTokens wins.+func (o *ChatOptions) CompletionBudget() int {+ if o == nil {+ return 0+ }+ if o.MaxCompletionTokens > 0 {+ return o.MaxCompletionTokens+ }+ return o.MaxTokens+}++// wireCompletionTokenField picks the Chat Completions JSON key for this+// provider+model, following the earendil-works/pi compat.maxTokensField+// pattern: one internal budget, exactly one outbound field.+//+// Default is max_completion_tokens (OpenAI Chat Completions, Azure, Ark),+// matching Pi. Only providers whose docs (or Pi's catalog) use the legacy+// name stay on max_tokens. WeKnora-only hosts that document max_tokens+// (LKEAP) are listed here too. Unknown OpenAI-compat hosts — including+// Aliyun DashScope — keep the default. GPT-5 / o-series always use+// max_completion_tokens.+func wireCompletionTokenField(name provider.ProviderName, model string) completionTokenField {+ if provider.IsOpenAIReasoningOrGPT5Model(model) {+ return completionTokenFieldMaxCompletionTokens+ }+ switch name {+ case provider.ProviderDeepSeek, // api-docs.deepseek.com: max_tokens only+ provider.ProviderZhipu, // open.bigmodel.cn: max_tokens only (Pi isZai)+ provider.ProviderSiliconFlow, // docs.siliconflow.com schema: max_tokens+ provider.ProviderMoonshot, // Pi useMaxTokens (moonshot.ai)+ provider.ProviderNvidia, // Pi useMaxTokens (NIM / vLLM)+ provider.ProviderGeneric, // self-hosted vLLM typically ignores the new field+ provider.ProviderGPUStack, // private vLLM-class runtime+ provider.ProviderLKEAP: // cloud.tencent.com/document/product/1772/115969: max_tokens only+ return completionTokenFieldMaxTokens+ default:+ return completionTokenFieldMaxCompletionTokens+ }+}++func applyCompletionBudget(req *openai.ChatCompletionRequest, budget int, field completionTokenField) {+ if req == nil || budget <= 0 {+ return+ }+ switch field {+ case completionTokenFieldMaxTokens:+ req.MaxTokens = budget+ default:+ req.MaxCompletionTokens = budget+ }+}diff --git a/internal/models/chat/completion_budget_test.go b/internal/models/chat/completion_budget_test.gonew file mode 100644index 000000000..90d2cc969--- /dev/null+++ b/internal/models/chat/completion_budget_test.go@@ -0,0 +1,177 @@+package chat++import (+ "context"+ "encoding/json"+ "testing"++ "github.com/Tencent/WeKnora/internal/models/provider"+ "github.com/stretchr/testify/assert"+ "github.com/stretchr/testify/require"+)++func TestChatOptionsCompletionBudget(t *testing.T) {+ assert.Zero(t, (*ChatOptions)(nil).CompletionBudget())+ assert.Zero(t, (&ChatOptions{}).CompletionBudget())+ assert.Equal(t, 128, (&ChatOptions{MaxTokens: 128}).CompletionBudget())+ assert.Equal(t, 256, (&ChatOptions{MaxCompletionTokens: 256}).CompletionBudget())+ assert.Equal(t, 256, (&ChatOptions{MaxTokens: 128, MaxCompletionTokens: 256}).CompletionBudget())+}++func TestWireCompletionTokenField(t *testing.T) {+ // Every AllProviders() name must be classified here so a new vendor cannot+ // silently inherit the default. Match Pi's polarity: default+ // max_completion_tokens; only documented max_tokens hosts (plus WeKnora+ // self-hosted / LKEAP, which Pi does not catalog) are exceptions.+ want := map[provider.ProviderName]completionTokenField{+ provider.ProviderDeepSeek: completionTokenFieldMaxTokens,+ provider.ProviderZhipu: completionTokenFieldMaxTokens,+ provider.ProviderSiliconFlow: completionTokenFieldMaxTokens,+ provider.ProviderMoonshot: completionTokenFieldMaxTokens,+ provider.ProviderNvidia: completionTokenFieldMaxTokens,+ provider.ProviderGeneric: completionTokenFieldMaxTokens,+ provider.ProviderGPUStack: completionTokenFieldMaxTokens,+ provider.ProviderLKEAP: completionTokenFieldMaxTokens,+ provider.ProviderOpenAI: completionTokenFieldMaxCompletionTokens,+ provider.ProviderAzureOpenAI: completionTokenFieldMaxCompletionTokens,+ provider.ProviderVolcengine: completionTokenFieldMaxCompletionTokens,+ provider.ProviderAliyun: completionTokenFieldMaxCompletionTokens,+ provider.ProviderLiteLLM: completionTokenFieldMaxCompletionTokens,+ provider.ProviderGemini: completionTokenFieldMaxCompletionTokens,+ provider.ProviderWeKnoraCloud: completionTokenFieldMaxCompletionTokens,+ provider.ProviderHunyuan: completionTokenFieldMaxCompletionTokens,+ provider.ProviderMiniMax: completionTokenFieldMaxCompletionTokens,+ provider.ProviderOpenRouter: completionTokenFieldMaxCompletionTokens,+ provider.ProviderRequesty: completionTokenFieldMaxCompletionTokens,+ provider.ProviderJina: completionTokenFieldMaxCompletionTokens,+ provider.ProviderMimo: completionTokenFieldMaxCompletionTokens,+ provider.ProviderModelScope: completionTokenFieldMaxCompletionTokens,+ provider.ProviderQianfan: completionTokenFieldMaxCompletionTokens,+ provider.ProviderQiniu: completionTokenFieldMaxCompletionTokens,+ provider.ProviderLongCat: completionTokenFieldMaxCompletionTokens,+ provider.ProviderNovita: completionTokenFieldMaxCompletionTokens,+ provider.ProviderAnthropic: completionTokenFieldMaxCompletionTokens,+ }+ require.Len(t, want, len(provider.AllProviders()), "classify every AllProviders() name")+ for _, name := range provider.AllProviders() {+ field, ok := want[name]+ require.True(t, ok, "classify %s in TestWireCompletionTokenField", name)+ assert.Equal(t, field, wireCompletionTokenField(name, "any"), string(name))+ }++ // GPT-5 / o-series always use the modern field, even on a max_tokens provider.+ assert.Equal(t, completionTokenFieldMaxCompletionTokens,+ wireCompletionTokenField(provider.ProviderGeneric, "gpt-5-mini"))+}++func TestBuildChatCompletionRequest_OneWireTokenField(t *testing.T) {+ messages := []Message{{Role: "user", Content: "hello"}}++ t.Run("volcengine both aliases send only max_completion_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderVolcengine), "doubao-seed-2-0-mini", nil)+ req := c.BuildChatCompletionRequest(messages, &ChatOptions{+ MaxTokens: 2048, MaxCompletionTokens: 4096,+ }, false)+ assert.Equal(t, 4096, req.MaxCompletionTokens)+ assert.Zero(t, req.MaxTokens)++ body, err := json.Marshal(req)+ require.NoError(t, err)+ assert.Contains(t, string(body), `"max_completion_tokens":4096`)+ assert.NotContains(t, string(body), `"max_tokens"`)+ })++ t.Run("deepseek sends max_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderDeepSeek), "deepseek-chat", nil)+ req := c.shapedRequest(messages, &ChatOptions{+ MaxTokens: 2048, MaxCompletionTokens: 4096,+ }, false)+ assert.Equal(t, 4096, req.MaxTokens)+ assert.Zero(t, req.MaxCompletionTokens)++ body, err := json.Marshal(req)+ require.NoError(t, err)+ assert.Contains(t, string(body), `"max_tokens":4096`)+ assert.NotContains(t, string(body), "max_completion_tokens")+ })++ t.Run("lkeap sends max_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderLKEAP), "deepseek-v3.1", nil)+ req := c.BuildChatCompletionRequest(messages, &ChatOptions{MaxCompletionTokens: 4096}, false)+ assert.Equal(t, 4096, req.MaxTokens)+ assert.Zero(t, req.MaxCompletionTokens)++ body, err := json.Marshal(req)+ require.NoError(t, err)+ assert.Contains(t, string(body), `"max_tokens":4096`)+ assert.NotContains(t, string(body), "max_completion_tokens")+ })++ t.Run("generic vLLM sends max_tokens from MaxTokens-only callers", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderGeneric), "qwen3", nil)+ req := c.BuildChatCompletionRequest(messages, &ChatOptions{MaxTokens: 2048}, false)+ assert.Equal(t, 2048, req.MaxTokens)+ assert.Zero(t, req.MaxCompletionTokens)+ })++ t.Run("aliyun dashscope keeps max_completion_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderAliyun), "qwen-plus", nil)+ req := c.BuildChatCompletionRequest(messages, &ChatOptions{MaxTokens: 2048}, false)+ assert.Zero(t, req.MaxTokens)+ assert.Equal(t, 2048, req.MaxCompletionTokens)+ })++ t.Run("openai gpt-4o sends max_completion_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderOpenAI), "gpt-4o", nil)+ req := c.BuildChatCompletionRequest(messages, &ChatOptions{MaxTokens: 128}, false)+ assert.Zero(t, req.MaxTokens)+ assert.Equal(t, 128, req.MaxCompletionTokens)+ })+}++func TestBuildOutbound_OneWireTokenField(t *testing.T) {+ msgs := []Message{{Role: "user", Content: "hello"}}++ t.Run("volcengine thinking keeps only max_completion_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderVolcengine), "doubao-seed-2-0-mini", nil)+ body, _, useRaw, err := c.buildOutbound(context.Background(), msgs, &ChatOptions{+ MaxTokens: 2048, MaxCompletionTokens: 4096, Thinking: ptrBool(true),+ }, true)+ require.NoError(t, err)+ require.True(t, useRaw)+ js := mustJSON(t, body)+ assert.Contains(t, js, `"thinking"`)+ assert.Contains(t, js, `"max_completion_tokens":4096`)+ assert.NotContains(t, js, `"max_tokens"`)+ })++ t.Run("lkeap thinking keeps only max_tokens", func(t *testing.T) {+ c := newOutboundChat(t, string(provider.ProviderLKEAP), "deepseek-v3.1", nil)+ body, _, useRaw, err := c.buildOutbound(context.Background(), msgs, &ChatOptions{+ MaxCompletionTokens: 4096, Thinking: ptrBool(false),+ }, true)+ require.NoError(t, err)+ require.True(t, useRaw)+ js := mustJSON(t, body)+ assert.Contains(t, js, `"thinking"`)+ assert.Contains(t, js, `"max_tokens":4096`)+ assert.NotContains(t, js, "max_completion_tokens")+ })+}++func TestOllamaBuildChatRequestUsesCompletionBudget(t *testing.T) {+ c := &OllamaChat{modelName: "llama3"}+ req := c.buildChatRequest(+ []Message{{Role: "user", Content: "hi"}},+ &ChatOptions{MaxCompletionTokens: 512},+ false,+ )+ assert.Equal(t, 512, req.Options["num_predict"])+}++func TestBuildLangfuseModelParamsUsesCompletionBudget(t *testing.T) {+ params := buildLangfuseModelParams(&ChatOptions{MaxTokens: 128, MaxCompletionTokens: 256})+ assert.Equal(t, 256, params["max_completion_tokens"])+ _, hasLegacy := params["max_tokens"]+ assert.False(t, hasLegacy)+}diff --git a/internal/models/chat/langfuse_wrapper.go b/internal/models/chat/langfuse_wrapper.goindex d1cf1abff..a012dd0b3 100644--- a/internal/models/chat/langfuse_wrapper.go+++ b/internal/models/chat/langfuse_wrapper.go@@ -193,11 +193,8 @@ func buildLangfuseModelParams(opts *ChatOptions) map[string]interface{} {if opts.TopP != 0 {params["top_p"] = opts.TopP}- if opts.MaxTokens > 0 {- params["max_tokens"] = opts.MaxTokens- }- if opts.MaxCompletionTokens > 0 {- params["max_completion_tokens"] = opts.MaxCompletionTokens+ if budget := opts.CompletionBudget(); budget > 0 {+ params["max_completion_tokens"] = budget}if opts.FrequencyPenalty != 0 {params["frequency_penalty"] = opts.FrequencyPenaltydiff --git a/internal/models/chat/llm_debug.go b/internal/models/chat/llm_debug.goindex 3987b7023..df6c12098 100644--- a/internal/models/chat/llm_debug.go+++ b/internal/models/chat/llm_debug.go@@ -57,11 +57,8 @@ func buildOptionsSection(opts *ChatOptions) string {if opts.TopP > 0 {parts = append(parts, fmt.Sprintf("TopP=%.2f", opts.TopP))}- if opts.MaxTokens > 0 {- parts = append(parts, fmt.Sprintf("MaxTokens=%d", opts.MaxTokens))- }- if opts.MaxCompletionTokens > 0 {- parts = append(parts, fmt.Sprintf("MaxCompletionTokens=%d", opts.MaxCompletionTokens))+ if budget := opts.CompletionBudget(); budget > 0 {+ parts = append(parts, fmt.Sprintf("CompletionBudget=%d", budget))}if opts.FrequencyPenalty > 0 {parts = append(parts, fmt.Sprintf("FrequencyPenalty=%.2f", opts.FrequencyPenalty))diff --git a/internal/models/chat/ollama.go b/internal/models/chat/ollama.goindex 44ba351cf..0b540c159 100644--- a/internal/models/chat/ollama.go+++ b/internal/models/chat/ollama.go@@ -103,8 +103,8 @@ func (c *OllamaChat) buildChatRequest(messages []Message, opts *ChatOptions, isSif opts.TopP > 0 {chatReq.Options["top_p"] = opts.TopP}- if opts.MaxTokens > 0 {- chatReq.Options["num_predict"] = opts.MaxTokens+ if budget := opts.CompletionBudget(); budget > 0 {+ chatReq.Options["num_predict"] = budget}if opts.Thinking != nil {chatReq.Think = &ollamaapi.ThinkValue{diff --git a/internal/models/chat/openai_request.go b/internal/models/chat/openai_request.goindex 85b680d5b..2f995ceae 100644--- a/internal/models/chat/openai_request.go+++ b/internal/models/chat/openai_request.go@@ -94,10 +94,11 @@ func (c *RemoteAPIChat) ConvertMessages(messages []Message) []openai.ChatComplet// BuildChatCompletionRequest 构建标准聊天请求参数(导出供子类使用)。//-// 这是一个不含任何 provider 特定逻辑的通用实现:所有采样参数(temperature /-// top_p / penalties)与 max_tokens 都按 opts 直接映射。供应商相关的特判-// (OpenAI o-series / GPT-5 改用 max_completion_tokens、Moonshot 固定温度等)-// 由对应的 providerAdapter.ShapeRequest 在事后施加,见 provider.go。+// 采样参数(temperature / top_p / penalties)按 opts 直接映射。完成预算经+// CompletionBudget 收成一个值,再按供应商只写入 max_tokens 或+// max_completion_tokens 之一(二者互斥,见 #3014)。其余供应商特判+// (o-series / GPT-5 采样参数、Moonshot 固定温度等)仍由+// providerAdapter.ShapeRequest 在事后施加,见 provider.go。func (c *RemoteAPIChat) BuildChatCompletionRequest(messages []Message, opts *ChatOptions, isStream bool,) openai.ChatCompletionRequest {@@ -126,12 +127,7 @@ func (c *RemoteAPIChat) BuildChatCompletionRequest(req.PresencePenalty = float32(opts.PresencePenalty)}- if opts.MaxTokens > 0 {- req.MaxTokens = opts.MaxTokens- }- if opts.MaxCompletionTokens > 0 {- req.MaxCompletionTokens = opts.MaxCompletionTokens- }+ applyCompletionBudget(&req, opts.CompletionBudget(), wireCompletionTokenField(c.provider, c.modelName))// 处理 Toolsif len(opts.Tools) > 0 {diff --git a/internal/models/chat/remote_api_test.go b/internal/models/chat/remote_api_test.goindex 865237397..a34e43be1 100644--- a/internal/models/chat/remote_api_test.go+++ b/internal/models/chat/remote_api_test.go@@ -188,8 +188,8 @@ func TestBuildChatCompletionRequest_GPT5MaxCompletionTokens(t *testing.T) {assert.EqualValues(t, 0, req.FrequencyPenalty, "frequency_penalty must be omitted")assert.EqualValues(t, 0, req.PresencePenalty, "presence_penalty must be omitted")} else {- assert.Equal(t, 128, req.MaxTokens)- assert.Equal(t, 0, req.MaxCompletionTokens)+ assert.Zero(t, req.MaxTokens, "non-reasoning OpenAI/Azure send max_completion_tokens only")+ assert.Equal(t, 128, req.MaxCompletionTokens)assert.InDelta(t, 0.7, req.Temperature, 1e-6)}})