fix(gateway): accept Claude Code max_tokens=1 probes for any model

The claude_code_only probe exemption only fired for models containing
"haiku". Claude Code also sends max_tokens=1 requests without a system
prompt to the model currently in use, e.g. when switching models or
refreshing the context usage, so on sonnet/opus groups those requests failed
the system-prompt check and were rejected as non-Claude-Code clients.

Recognise the probe from the request body itself (max_tokens == 1) once the
User-Agent check has passed, regardless of model, and carry max_tokens into
the ParsedRequest projection so the hot path sees it too. The haiku context
flag and the local haiku interception are unchanged; non-haiku probes reach
the pool as ordinary one-token requests. One output token offers nothing to
a client that already forges the CLI User-Agent, so this widens no attack
surface.

Fixes #6591

backend/internal/handler/gateway_helper.go

@@ -89,6 +89,10 @@ func claudeCodeBodyMapFromParsedRequest(parsedReq *service.ParsedRequest) map[st
 	bodyMap := map[string]any{
 		"model": parsedReq.Model,
 	}
+	// 探测识别（max_tokens=1）需要看到该字段，复用已解析请求时一并带上。
+	if parsedReq.MaxTokens > 0 {
+		bodyMap["max_tokens"] = parsedReq.MaxTokens
+	}
 	if parsedReq.HasSystem {
 		if system, ok := parsedReq.SystemValue(); ok {
 			bodyMap["system"] = system

backend/internal/handler/gateway_helper_hotpath_test.go

@@ -500,3 +500,19 @@ type helperConcurrencyCacheStubWithError struct {
 func (s *helperConcurrencyCacheStubWithError) AcquireAccountSlot(ctx context.Context, accountID int64, maxConcurrency int, requestID string) (bool, error) {
 	return false, s.err
 }
+
+func TestSetClaudeCodeClientContext_ParsedRequestProbeWithoutSystemPrompt(t *testing.T) {
+	c, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
+	c.Request.Header.Set("User-Agent", "claude-cli/2.1.260 (external, cli)")
+
+	// The hot path reuses ParsedRequest instead of re-parsing the body; the probe
+	// marker must survive that projection or the exemption only works cold.
+	parsed := &service.ParsedRequest{Model: "claude-sonnet-4-5", MaxTokens: 1}
+	SetClaudeCodeClientContext(c, nil, parsed)
+	require.True(t, service.IsClaudeCodeClient(c.Request.Context()))
+
+	c2, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
+	c2.Request.Header.Set("User-Agent", "claude-cli/2.1.260 (external, cli)")
+	SetClaudeCodeClientContext(c2, nil, &service.ParsedRequest{Model: "claude-sonnet-4-5", MaxTokens: 64})
+	require.False(t, service.IsClaudeCodeClient(c2.Request.Context()))
+}

backend/internal/service/claude_code_validator.go

@@ -104,6 +104,12 @@ func (v *ClaudeCodeValidator) Validate(r *http.Request, body map[string]any) boo
 	if isMaxTokensOneHaiku, ok := IsMaxTokensOneHaikuRequestFromContext(r.Context()); ok && isMaxTokensOneHaiku {
 		return true // 绕过 system prompt 检查，UA 已在 Step 1 验证
 	}
+	// 探测请求并不总是打到 haiku：CLI 切换模型、刷新上下文用量时会向当前模型发
+	// max_tokens=1 的轻量请求，同样不携带 system。UA 已过 Step 1，且 1 个输出 token
+	// 对滥用者没有价值，故按请求体放行，不再限定模型名。
+	if isMaxTokensOneBody(body) {
+		return true
+	}
 
 	// Step 4: messages 路径，进行严格验证
 
@@ -154,6 +160,24 @@ func isMessagesCountTokensPath(path string) bool {
 	return strings.HasSuffix(path, "/messages/count_tokens")
 }
 
+// isMaxTokensOneBody 判断请求体是否显式声明 max_tokens=1。
+// 兼容 JSON 反序列化出的 float64 与 ParsedRequest 复用时的 int。
+func isMaxTokensOneBody(body map[string]any) bool {
+	if body == nil {
+		return false
+	}
+	switch v := body["max_tokens"].(type) {
+	case float64:
+		return v == 1
+	case int:
+		return v == 1
+	case int64:
+		return v == 1
+	default:
+		return false
+	}
+}
+
 // hasClaudeCodeSystemPrompt 检查请求是否包含 Claude Code 系统提示词
 // 使用字符串相似度匹配（Dice coefficient）
 func (v *ClaudeCodeValidator) hasClaudeCodeSystemPrompt(body map[string]any) bool {

backend/internal/service/claude_code_validator_test.go

@@ -45,9 +45,10 @@ func TestClaudeCodeValidator_MessagesWithoutProbeStillNeedStrictValidation(t *te
 	req := httptest.NewRequest(http.MethodPost, "http://example.com/v1/messages", nil)
 	req.Header.Set("User-Agent", "claude-cli/1.2.3 (darwin; arm64)")
 
+	// max_tokens=2 不是探测请求：没有 system prompt 仍走严格校验并被拒。
 	ok := validator.Validate(req, map[string]any{
 		"model":      "claude-haiku-4-5",
-		"max_tokens": 1,
+		"max_tokens": 2,
 	})
 	require.False(t, ok)
 }
@@ -574,3 +575,26 @@ func TestSetGetClaudeCodeVersion(t *testing.T) {
 	ctx = SetClaudeCodeVersion(ctx, "2.1.63")
 	require.Equal(t, "2.1.63", GetClaudeCodeVersion(ctx))
 }
+
+func TestClaudeCodeValidator_MaxTokensOneProbeIsNotLimitedToHaiku(t *testing.T) {
+	validator := NewClaudeCodeValidator()
+
+	for _, model := range []string{"claude-sonnet-4-5", "claude-opus-4-1", "claude-haiku-4-5"} {
+		t.Run(model, func(t *testing.T) {
+			req := httptest.NewRequest(http.MethodPost, "http://example.com/v1/messages", nil)
+			req.Header.Set("User-Agent", "claude-cli/2.1.260 (external, cli)")
+			// No context flag, no system prompt, no extra headers: the body alone marks the probe.
+			for _, mt := range []any{float64(1), 1} {
+				require.True(t, validator.Validate(req, map[string]any{"model": model, "max_tokens": mt}), "max_tokens=%v (%T)", mt, mt)
+			}
+		})
+	}
+}
+
+func TestClaudeCodeValidator_MaxTokensOneProbeStillRequiresClaudeCodeUA(t *testing.T) {
+	validator := NewClaudeCodeValidator()
+	req := httptest.NewRequest(http.MethodPost, "http://example.com/v1/messages", nil)
+	req.Header.Set("User-Agent", "python-requests/2.32")
+
+	require.False(t, validator.Validate(req, map[string]any{"model": "claude-sonnet-4-5", "max_tokens": 1}))
+}