feat(ollama): 对 Ollama Cloud 账号 clamp max_tokens 上限

backend/internal/service/openai_gateway_ollama_cloud_cc_reasoning.go

@@ -51,7 +51,8 @@ func applyOllamaCloudRawChatCompletionsRequest(account *Account, body []byte) []
 	if !isOllamaCloudRawChatCompletionsAccount(account) || len(body) == 0 {
 		return body
 	}
-	return normalizeOllamaCloudChatCompletionsRequest(body)
+	body = normalizeOllamaCloudChatCompletionsRequest(body)
+	return clampOllamaCloudMaxTokens(account, body)
 }
 
 func applyOllamaCloudRawChatCompletionsResponse(account *Account, body []byte) []byte {

backend/internal/service/openai_gateway_ollama_cloud_max_tokens.go

@@ -0,0 +1,83 @@
+package service
+
+import (
+	"encoding/json"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/logger"
+	"github.com/tidwall/gjson"
+	"github.com/tidwall/sjson"
+	"go.uber.org/zap"
+)
+
+// OllamaCloudMaxTokensCapExtraKey 是账号 extra 中的可选配置键，表示该 Ollama Cloud
+// 账号输出 token 的 provider 级硬上限。用户可通过 admin 账号更新 API 的 extra 字段
+// 设置，覆盖默认值 ollamaCloudDefaultMaxTokensCap；0 或负数表示显式禁用 clamp。
+const OllamaCloudMaxTokensCapExtraKey = "ollama_max_tokens_cap"
+
+// ollamaCloudDefaultMaxTokensCap 是 Ollama Cloud 对输出 token 数的 provider 级硬上限
+// （约 65535），max_tokens 超过该值会被上游直接 400 拒绝；该上限与模型无关，不做模型过滤。
+const ollamaCloudDefaultMaxTokensCap = 65535
+
+// 本文件的 clampOllamaCloudMaxTokens 被
+// applyOllamaCloudRawChatCompletionsRequest（openai_gateway_ollama_cloud_cc_reasoning.go）
+// 调用，账号检测（isOllamaCloudRawChatCompletionsAccount）由调用方完成，此处不再重复判断。
+
+// ollamaCloudMaxTokensCap 返回账号配置的 max_tokens 上限。账号为 nil 或 extra 中
+// 无该键时返回默认值；键值为数值类型（float64/int64/int/json.Number）时返回其整数
+// 值（0 或负数表示显式禁用 clamp）；其它类型回退默认值。
+func ollamaCloudMaxTokensCap(account *Account) int64 {
+	if account == nil || account.Extra == nil {
+		return ollamaCloudDefaultMaxTokensCap
+	}
+	value, ok := account.Extra[OllamaCloudMaxTokensCapExtraKey]
+	if !ok {
+		return ollamaCloudDefaultMaxTokensCap
+	}
+	switch number := value.(type) {
+	case float64:
+		return int64(number)
+	case int64:
+		return number
+	case int:
+		return int64(number)
+	case json.Number:
+		parsed, err := number.Int64()
+		if err != nil {
+			return ollamaCloudDefaultMaxTokensCap
+		}
+		return parsed
+	default:
+		return ollamaCloudDefaultMaxTokensCap
+	}
+}
+
+// clampOllamaCloudMaxTokens 把 body 中超过 cap 的 max_tokens / max_completion_tokens
+// 单向压到 cap。cap <= 0 或 body 不是合法 JSON 时原样返回；sjson 出错时返回原始 body。
+// 有任一字段被 clamp 时记录一条 Debug 日志。
+func clampOllamaCloudMaxTokens(account *Account, body []byte) []byte {
+	cap := ollamaCloudMaxTokensCap(account)
+	if cap <= 0 || !gjson.ValidBytes(body) {
+		return body
+	}
+	clamped := false
+	out := body
+	for _, key := range []string{"max_tokens", "max_completion_tokens"} {
+		result := gjson.GetBytes(out, key)
+		if !result.Exists() || result.Type != gjson.Number || result.Int() <= cap {
+			continue
+		}
+		updated, err := sjson.SetBytes(out, key, cap)
+		if err != nil {
+			return body
+		}
+		out = updated
+		clamped = true
+	}
+	if clamped && account != nil {
+		logger.L().Debug("openai chat_completions raw: clamped max_tokens for ollama cloud account",
+			zap.Int64("account_id", account.ID),
+			zap.Int64("cap", cap),
+		)
+	}
+	return out
+}

backend/internal/service/openai_gateway_ollama_cloud_max_tokens_test.go

@@ -0,0 +1,160 @@
+//go:build unit
+
+package service
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/Wei-Shaw/sub2api/internal/pkg/openai_compat"
+	"github.com/stretchr/testify/require"
+)
+
+// ollamaMaxTokensCapTestAccount 构造带自定义 cap 的 Ollama Cloud usage 账号。
+func ollamaMaxTokensCapTestAccount(id int64, cap any) *Account {
+	account := ollamaUsageAccount(id)
+	account.Extra[OllamaCloudMaxTokensCapExtraKey] = cap
+	return account
+}
+
+func TestOllamaCloudMaxTokensClamp(t *testing.T) {
+	ollama := ollamaUsageAccount(101)
+
+	tests := []struct {
+		name    string
+		account *Account
+		body    string
+		want    string
+		raw     bool // want 非法 JSON 时按原始字节比较
+	}{
+		{
+			name:    "max_tokens above default cap is clamped",
+			account: ollama,
+			body:    `{"model":"gpt-oss:120b-cloud","max_tokens":70000}`,
+			want:    `{"model":"gpt-oss:120b-cloud","max_tokens":65535}`,
+		},
+		{
+			name:    "max_completion_tokens above default cap is clamped",
+			account: ollama,
+			body:    `{"model":"gpt-oss:120b-cloud","max_completion_tokens":131072}`,
+			want:    `{"model":"gpt-oss:120b-cloud","max_completion_tokens":65535}`,
+		},
+		{
+			name:    "both fields above cap are clamped",
+			account: ollama,
+			body:    `{"model":"m","max_tokens":80000,"max_completion_tokens":90000}`,
+			want:    `{"model":"m","max_tokens":65535,"max_completion_tokens":65535}`,
+		},
+		{
+			name:    "values at or below default cap are kept",
+			account: ollama,
+			body:    `{"model":"m","max_tokens":65535,"max_completion_tokens":4096}`,
+			want:    `{"model":"m","max_tokens":65535,"max_completion_tokens":4096}`,
+		},
+		{
+			name:    "custom extra cap is applied",
+			account: ollamaMaxTokensCapTestAccount(102, 32768),
+			body:    `{"model":"m","max_tokens":50000}`,
+			want:    `{"model":"m","max_tokens":32768}`,
+		},
+		{
+			name:    "extra cap zero disables clamping",
+			account: ollamaMaxTokensCapTestAccount(103, 0),
+			body:    `{"model":"m","max_tokens":50000}`,
+			want:    `{"model":"m","max_tokens":50000}`,
+		},
+		{
+			name:    "non-numeric extra cap falls back to default",
+			account: ollamaMaxTokensCapTestAccount(104, "abc"),
+			body:    `{"model":"m","max_tokens":100000}`,
+			want:    `{"model":"m","max_tokens":65535}`,
+		},
+		{
+			name:    "invalid json is left untouched",
+			account: ollama,
+			body:    `{"model":"m","max_tokens":`,
+			want:    `{"model":"m","max_tokens":`,
+			raw:     true,
+		},
+		{
+			name:    "non-integer max_tokens is left untouched",
+			account: ollama,
+			body:    `{"model":"m","max_tokens":1.5}`,
+			want:    `{"model":"m","max_tokens":1.5}`,
+		},
+		{
+			name:    "missing max_tokens is left untouched",
+			account: ollama,
+			body:    `{"model":"m"}`,
+			want:    `{"model":"m"}`,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got := clampOllamaCloudMaxTokens(test.account, []byte(test.body))
+			if test.raw {
+				require.Equal(t, test.want, string(got))
+				return
+			}
+			require.JSONEq(t, test.want, string(got))
+		})
+	}
+}
+
+func TestOllamaCloudMaxTokensCap(t *testing.T) {
+	require.Equal(t, int64(65535), ollamaCloudMaxTokensCap(nil))
+	require.Equal(t, int64(65535), ollamaCloudMaxTokensCap(ollamaUsageAccount(201)))
+
+	tests := []struct {
+		name string
+		cap  any
+		want int64
+	}{
+		{"float64", float64(32768), 32768},
+		{"int", 40000, 40000},
+		{"int64", int64(50000), 50000},
+		{"json.Number", json.Number("60000"), 60000},
+		{"json.Number invalid", json.Number("abc"), 65535},
+		{"zero disables", 0, 0},
+		{"negative disables", int64(-1), -1},
+		{"string falls back", "abc", 65535},
+		{"bool falls back", true, 65535},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			account := ollamaMaxTokensCapTestAccount(202, test.cap)
+			require.Equal(t, test.want, ollamaCloudMaxTokensCap(account))
+		})
+	}
+}
+
+// TestApplyOllamaCloudRawChatCompletionsRequestClampsMaxTokens 验证 max_tokens clamp
+// 已接入组合钩子 applyOllamaCloudRawChatCompletionsRequest，并遵循该钩子的账号判定门槛
+// （isOllamaCloudRawChatCompletionsAccount：platform openai + type apikey +
+// force_chat_completions + ollama.com 或 Ollama usage extra）。
+func TestApplyOllamaCloudRawChatCompletionsRequestClampsMaxTokens(t *testing.T) {
+	body := []byte(`{"model":"deepseek-chat","max_tokens":100000}`)
+
+	// Ollama Cloud 账号（ollama.com + force_chat_completions）→ clamp 到 65535。
+	ollama := ollamaCloudRawChatCompletionsTestAccount()
+	require.JSONEq(t, `{"model":"deepseek-chat","max_tokens":65535}`,
+		string(applyOllamaCloudRawChatCompletionsRequest(ollama, body)))
+
+	// 官方 DeepSeek（api.deepseek.com + force_chat_completions）→ 字节级不变。
+	official := rawChatCompletionsTestAccount()
+	official.Credentials["base_url"] = "https://api.deepseek.com"
+	official.Extra = map[string]any{
+		openai_compat.ExtraKeyResponsesMode: string(openai_compat.ResponsesSupportModeForceChatCompletions),
+	}
+	require.Equal(t, body, applyOllamaCloudRawChatCompletionsRequest(official, body))
+
+	// ollama.com 但无 force_chat_completions（Extra 缺键）→ 不通过钩子判定门槛，字节级不变。
+	noForce := ollamaCloudRawChatCompletionsTestAccount()
+	noForce.Extra = nil
+	require.Equal(t, body, applyOllamaCloudRawChatCompletionsRequest(noForce, body))
+
+	// 空 body → 原样返回。
+	require.Equal(t, []byte(nil), applyOllamaCloudRawChatCompletionsRequest(ollama, nil))
+	require.Equal(t, []byte{}, applyOllamaCloudRawChatCompletionsRequest(ollama, []byte{}))
+}